Is Anthropic acting responsibly while advancing toward recursive self-improvement?
After leaving Anthropic, Jacob Coxon said Anthropic and OpenAI were irresponsibly racing toward self-improving superintelligence, while praising Anthropic's current practices. Responses dispute catastrophic-risk forecasts, working inside AI labs, and whether governments should pause or regulate development.
BackgroundWhat is frontier AI safety?
Case period:
Published by The Dispute Index editorial teamPublished Updated
Overview
Jacob Coxon resigned from Anthropic and said that Anthropic and OpenAI are acting irresponsibly by racing toward self-improving superintelligence. In a fuller interview, he praised Anthropic's current practices and denied that it was already cutting corners; his criticism concerns the competitive race and the pressures he expects it to create. Responses dispute the evidence and timing of catastrophic-risk forecasts, whether continued work inside AI labs can be justified, and whether governments should pause development or regulate its continuation. The record distinguishes these disagreements from arguments about the rhetoric of public warnings.
Background: What is frontier AI safety?
Read the full background
Coxon's complete resignation thread is the primary record. His WIRED interview supplies essential context: he describes Anthropic as the most responsible major lab in his experience and says it is not currently cutting corners, while arguing that no private company should decide the collective risk of the race. His warning is about a future competitive dynamic, not a claim that today's consumer models are already causing human extinction.
Hubinger acknowledges severe future risk, gives a personal estimate above 10% within a decade, and says Anthropic lacks a demonstrated alignment plan for superintelligence. He also says present-model risk is low and Anthropic is trying its best. That last point is not itself a disagreement with Coxon. Marks and Wang discuss reducing risks through continued work inside a lab; Elmore directly challenges Wang's decision to stay, and Abaluck defends continued capabilities work only under demanding conditions.
Lambert challenges the evidence for the numerical risk characterization; Marcus disputes the near-term timeline while criticizing lab reasoning. Barak expresses doubt about universal human extinction but still urges safety and coordination. Sanders explicitly endorses Coxon and proposes a ban and pause. Jones favors regulated innovation and an international treaty; Schulman calls for a joint industry pacing proposal. Kim, Lorenz, Burgis and Krishnan raise distinct objections about warning rhetoric, evidence, moral framing or whose views represent the industry. These are not interchangeable defenses of corporate conduct.
Some positions are replies to Coxon, others respond to Hubinger or Wang, and some are contemporaneous commentary without a preserved direct-reply relationship. Each entry identifies that scope. Silence, refusal to comment, and institutional endorsement are not inferred. Employees' personal statements are not corporate responses. Anthropic's June 4 announcement and August 31 safety measures predate the resignation and remain background, not additional people or direct replies. Forecasts and allegations are attributed to their speakers.
People in this case
Jacob CoxonDirectly involved
Evan HubingerDirectly involved
Samuel MarksDirectly involved
Timeline
20 timeline entries on this page. Dates: August 31, 2026 to October 5, 2026
August 2026
Anthropic describes post-incident safeguards and pacing
Source release
Anthropic said it paused some external cyber evaluations and higher-risk reinforcement-learning environments, added monitoring and containment measures, and endorsed coordinated pacing across the industry.
[01]Anthropic
Source excerpt
“Within a company, pacing means a series of decisions that prioritize safety over speed when the two are in tension.”
Anthropic’s official account of recent incidents and its response. The company says it paused some external cyber evaluations, paused higher-risk reinforcement-learning environments for several weeks, made changes, and supports lawful, verifiable coordinated pacing across the industry.
Anthropic: Improving our alignment and security efforts · AnthropicPacing, paused evaluations, and alignment-assessment sections[01]Anthropic
Source excerpt
“Within a company, pacing means a series of decisions that prioritize safety over speed when the two are in tension.”
Anthropic’s official account of recent incidents and its response. The company says it paused some external cyber evaluations, paused higher-risk reinforcement-learning environments for several weeks, made changes, and supports lawful, verifiable coordinated pacing across the industry.
Anthropic: Improving our alignment and security efforts · AnthropicPacing, paused evaluations, and alignment-assessment sections
September 2026
Samuel Marks supports Coxon’s concern in a personal-capacity statement
Reaction
Marks said staff at AI developers want to slow down to improve safety and described unresolved robust alignment alongside his rationale for continuing safety work.
[02]Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risks
Source excerpt
“Many AI developer staff desperately want to slow down to figure out how to build AI more safely.”
Primary statement by an Anthropic safety researcher in his personal capacity. Marks supports Coxon’s central risk and coordination concerns while explaining why he continues safety work at Anthropic.
Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risksX post, personal capacity disclaimerJacob Coxon resigns and calls the frontier AI race irresponsible
Claim
Coxon announced his resignation from Anthropic and argued that Anthropic and OpenAI were racing toward self-improving superintelligence without acting responsibly.
[03]Jacob Coxon resigns from Anthropic and criticizes the frontier AI race
Source excerpt
“Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”
Primary record of Coxon’s resignation statement. He argues that Anthropic understands the stakes but is structurally locked into a race, and calls for coordination that could include a temporary capability pause.
Jacob Coxon resigns from Anthropic and criticizes the frontier AI raceSeven-post X threadEvan Hubinger gives a partial defense and concedes unresolved alignment
Reaction
Hubinger said Anthropic was trying its best but did not yet have a plan to solve alignment for superintelligence or a clear path to doing so.
[04]Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s efforts
Source excerpt
“I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”
Primary response by Anthropic’s Alignment Science Lead. Hubinger agrees with Coxon about the severity of potential future risk, credits Anthropic’s effort, and explicitly concedes the lack of a demonstrated alignment solution. His follow-up distinguishes present-model risk from future superintelligence risk.
Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s effortsX response and follow-upCoxon’s full interview distinguishes current safeguards from future race pressure
Source release
In WIRED's full interview, Coxon praises Anthropic's current practices and denies it is already cutting corners. He argues that competitive pressure could force future compromises and that the decision should not rest with a private company.
[05]Jacob Coxon’s full WIRED interview: resignation, current safeguards and future race pressures
Original full interview. Coxon distinguishes praise for Anthropic's current practices from his objection to future competitive pressure. He explicitly denies that Anthropic is already cutting corners. Includes discussion of both OpenAI and Anthropic; not merely the viral resignation excerpt.
Gallego requests a bipartisan Senate AI oversight committee
Ruben Gallego wrote to Senate leaders seeking a select committee with subpoena authority and a mandate to develop legislative recommendations. His letter cites the recent researcher resignation, and his office identifies Coxon. This records a request for the next Congress, not an established committee or a determination about lab misconduct.
[06]Gallego requests a Senate AI oversight committee after researcher warnings
Primary letter requests a bipartisan committee for the next Congress. Its official release explicitly names Coxon. This is a proposed oversight mechanism, not enacted legislation or a finding that either lab acted unlawfully.
Gallego requests a Senate AI oversight committee after researcher warningsSeptember 9 signed letter, pages1–3; resignation paragraph on page1 and committee mandate on pages2–3[07]Gallego links his proposed Senate AI committee to Coxon's resignation
Official Senate release identifies Coxon as the researcher whose resignation informed Gallego's request. Read with the separately cited signed letter. The request concerns a committee in the next Congress; no committee is established by this release.
Gallego links his proposed Senate AI committee to Coxon's resignationOpening release paragraphs naming Jacob Coxon and introducing the letter to Senate leadersAnthropic releases September threat-intelligence report
Source release
Anthropic said its September report described operations it identified and disrupted between December 2025 and August 2026, along with safeguards it said it strengthened. It said the covered cases did not involve Fable or Mythos except one illicit-distillation case. The report records Anthropic’s own account and does not independently resolve the case’s broader questions about AI-company conduct.
[08]Anthropic
Source excerpt
Anthropic said the report describes operations it identified and disrupted over an eight-month period.
An official Anthropic report. It describes the company’s own assessment of operations and safeguards; it is not independent verification and does not directly resolve the broader authorship-and-credit dispute in this case.
Anthropic’s September 2026 Threat Intelligence Report · AnthropicOverview and scope (web lines 41–47, 51–54).OpenAI researchers publish personal calls for slower AI development
Reaction
Steele supports slowing AI development and explicitly identifies her statement as personal despite her OpenAI employment. This is not an OpenAI institutional position. No conclusion is drawn about the company's response from Steele's employment. Wang warns about rapid movement toward recursive self-improvement and supports a petition asking the United States government to pace AI development. No direct reply relationship to Coxon is inferred. Wang's estimate of petition participation is not adopted as a verified count, and this is not an OpenAI statement. This entry preserves contextual commentary without classifying it as a direct stance on the case's central characterization. The timestamp marks Steele's post; Wang posted at 00:11:05 UTC on September 10, about five minutes later.
[09]Julie Steele
Source excerpt
I work at OpenAI. In my personal capacity, I also think we need to slow down.
Original post by the AI safety researcher, September 9 evening Eastern / September 10 UTC. The speaker explicitly identifies a personal capacity, not an OpenAI policy statement.
[10]Jasmine Wang
Source excerpt
It's hard to overstate how dangerous speeding towards RSI is.
Original personal post, September 9 evening Eastern / September 10 UTC. Wang says she signed a pacing petition. Her estimate of the share of lab employees signing is not independently established here. She is Jasmine Wang, not Anna Wang; no direct reply to Coxon is inferred.
Coxon distinguishes present AI systems from future risks on CNN
Reaction
In a September 9 CNN interview, Jacob Coxon distinguished present-day models from the future recursive self-improvement he warned about. He also said company leaders genuinely wanted regulation. This is additional primary context, not a retraction or a newly changed position. The displayed time marks the start of the interview segment at 20:35 Eastern, not an exact quotation timestamp.
[11]Jacob Coxon: CNN interview on present and future AI risks
Separates present models from future recursive self-improvement and says company leaders genuinely want regulation. This adds direct evidence/context rather than retracting his warning. Claims about autonomous cyber incidents and math results require independent checking if extracted. Segment marker20:35ET Sep9; not exact utterance time Full original CNN rush transcript; original date and speaker verified.
Jacob Coxon: CNN interview on present and future AI risksInterview from20:35ET through before20:46:25ET; present-risk quote before20:40, regulation comments after20:40Ward discusses AI competition and outside intervention
Reaction
Jacob Ward described Coxon's and other researchers' concern that competitive pressure prevents AI companies from slowing development. He argued that outside intervention was missing. This contextual interview does not constitute opposition to Coxon's warning. CNN places the passage after 01:15:24 and before 01:23:28 Eastern; the displayed time is the start of that interval.
[12]Jacob Ward: CNN interview on AI competition and regulation
Jacob Ward explains Coxon's and other researchers' concern that competitive pressure prevents labs stepping away, and argues outside intervention is missing. This does not challenge Coxon's warning. The date field uses the start of the transcript interval, 01:15:24 Eastern on September 10, not an exact utterance or webpage publication time. The passage ends before 01:23:28 Eastern. Original CNN rush transcript, subject to broadcaster revisions.
Jacob Ward: CNN interview on AI competition and regulationWard interview after01:15:24ET and before01:23:28ETCNBC reports company responses to renewed AI-safety warnings
Reaction
CNBC reports a general safety defense by Anthropic, a separate Anthropic refusal about Sacks's IPO post, and an OpenAI declined-comment response to employee warnings.
[13]CNBC reports Anthropic's safety response and two narrowly scoped declined-comment requests
Source excerpt
both enormous benefits and unprecedented risks / Anthropic declined to comment on the post. / OpenAI declined to comment when approached by CNBC
CNBC attributes a safety defense to an Anthropic spokesperson, separately says Anthropic declined a request about Sacks's post, and says OpenAI declined its employee-warning inquiry while noting recent blog posts. The two refusal excerpts are the reporter's wording, not verbatim corporate speech.
Coxon expands on future-risk scenarios in a CBS interview
Source release
CBS published Jo Ling Kent's interview with Jacob Coxon, in which he described possible future physical-world and biological risks while reiterating that current AI products are safe for ordinary use. The interview adds context to his warning rather than retracting it or establishing that advanced AI has already taken control. The article also preserves a separate Anthropic safety defense consistent with the company's previously recorded response.
[14]Jacob Coxon explains future AI-risk scenarios in a CBS News interview
Source excerpt
Now, imagine the AI refuses to turn on your light,
CBS's written account of Jo Ling Kent's September 10 interview with Coxon. He describes hypothetical advanced-AI risks while explicitly distinguishing current products as safe for ordinary use. The article includes a separate Anthropic spokesperson defense substantially overlapping the company response already represented here; it is not counted as a new stance. The interview is additional context, not a retraction or proof of the future scenarios.
Jacob Coxon explains future AI-risk scenarios in a CBS News interviewOpening interview paragraphs and the paragraph beginning 'Notably, Coxon said that current AI platforms'; corporate response follows under 'Anthropic defends safety of its AI'Former Anthropic safety-team member calls for external accountability
Reaction
Joe Benton, who said he had recently left Anthropic’s safety team, wrote that frontier-AI companies are racing to build recursively self-improving systems and said he would join METR, an independent evaluator. The account is his personal assessment and does not independently establish company practices.
[15]Joe Benton — Why I left Anthropic’s safety team to hold AI companies accountable
Source excerpt
Frontier AI companies are racing to build AI systems that can recursively self-improve.
Personal post by a former Anthropic safety-team member describing his assessment of frontier-AI incentives and his move to METR. It is attributed commentary from a former employee, not independent verification of company practices.
Amodei publishes a proposal to pace frontier AI development
Source release
Anthropic CEO Dario Amodei published an essay calling for frontier-AI development to be paced so safety measures can catch up. He said Anthropic would give third-party evaluators permanent employee-level access to its systems; the statement is a commitment, not evidence of completed implementation.
[16]Dario Amodei: We Must Pace the Frontier
Source excerpt
Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.
Direct public essay by Anthropic's CEO. It sets out an industry-pacing proposal and says Anthropic is committing to give embedded third-party evaluators ongoing, employee-like access; the statement is a commitment, not evidence that the measures have been implemented.
Dario Amodei: We Must Pace the FrontierEssay sections addressing recursive self-improvement and permanent third-party evaluatorsAltman says OpenAI will use independent evaluators with employee-like access
Reaction
According to Associated Press reporting, OpenAI CEO Sam Altman agreed that frontier AI should be paced and said OpenAI would also use independent evaluators with employee-like access. The statement is an attributed commitment, not evidence of implementation.
[17]Anthropic CEO Dario Amodei says AI industry needs to give safety measures time to catch up
Source excerpt
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.
Associated Press reporting on Amodei’s proposal and Altman’s public response.
Anthropic CEO Dario Amodei says AI industry needs to give safety measures time to catch upCoverage of Sam Altman’s public response to Amodei’s proposalTrump says possible AI guardrails must be weighed against competition
Reaction
Associated Press reported that Trump said possible AI guardrails could be used but dismissed unspecified warnings and emphasized maintaining the U.S. edge over China. He did not announce a particular rule, pace requirement, or determination about Anthropic’s conduct.
[18]Trump comments on AI guardrails, competition and development pace
Source excerpt
“We can put guardrails … But whoever wins with AI wins.”
Associated Press reports Trump’s response to questions about slowing or regulating AI shortly after Dario Amodei’s pacing proposal. The quoted remarks acknowledge possible guardrails while emphasizing U.S. competition with China; they do not specify a rule, implementation mechanism, or a direct assessment of Anthropic’s conduct.
Trump comments on AI guardrails, competition and development paceAP lines 1785–1790: Trump’s direct remarks to reporters during his September 13 trip to Ireland; surrounding context identifies the question about slowing or regulating AI.Amodei says future Anthropic releases require testing and independent evaluation
Reaction
Amodei said each new generation of Anthropic models should be tested and independently evaluated, while rejecting a complete ban and backing federal oversight. The interview follows Coxon’s public warning about competitive pressure and AI risk.
[19]Dario Amodei discusses AI testing and governance with CBS News
CBS News interview with Anthropic chief executive Dario Amodei, broadcast September 13 and updated on the site September 14. He supports testing and oversight for increasingly capable models while rejecting a complete ban.
Dario Amodei discusses AI testing and governance with CBS NewsInterview transcript, statements on testing, a kill switch, and joint governancePhilip DeFranco covers Coxon's warning and subsequent AI safety proposals
Source release
In the opening segment of his September 14 episode, Philip DeFranco reports on Coxon's resignation and warning, Amodei's safety proposals, and industry responses. This records the release of additional coverage, not a new finding about AI risk or an independent DeFranco position on Anthropic's conduct. His separate criticism of Washington's policy response is recorded in the frontier-AI safeguards case.
[20]Philip DeFranco
Contextual reporting in DeFranco's September 14 episode about Coxon's resignation, Amodei's safety proposals and industry responses. This source is supplementary coverage only, not a new DeFranco position on Anthropic's responsibility and not independent corroboration of extinction forecasts. Coxon's original thread and full WIRED interview remain the controlling evidence for his position, including his praise for Anthropic's current practices. This coverage is linked only to a source-release timeline entry, not appended as proof of the full Coxon claim. DeFranco's separate policy judgment at 05:37–06:06 belongs in the Washington safeguards case. The full timestamped YouTube transcript was reviewed; it contains transcription errors, so no new exact third-party quotation or technical claim is extracted here.
Philip DeFranco covers Coxon's warning and subsequent frontier-AI safety proposals · Philip DeFranco00:27–04:15; Coxon resignation and warning at 00:30–00:42Anthropic publishes metrics on AI-led R&D and safety oversight
Source release
Anthropic published a measurement post reporting that Claude was not fully autonomous in any measured subset of its AI R&D work, alongside an index of AI-led R&D, agent-oversight measures, and compute-allocation figures. The post says Anthropic plans embedded external evaluators and frames the disclosures as relevant to proposals to pace frontier AI development.
[21]Anthropic publishes metrics on AI-led R&D and safety oversight
Source excerpt
“Claude is not operating fully autonomously for any measured subset of AI R&D work.”
Anthropic’s public measurement post says its systems are automating more of model development while reporting that Claude was not fully autonomous in the measured subsets. It presents an R&D Automation Index, agent-oversight measures, compute-allocation figures, and a plan for embedded external evaluators in the context of calls to pace frontier AI.
Anthropic publishes metrics on AI-led R&D and safety oversightOpening; “What we found” sections on AI-led R&D, oversight, and compute allocation.
October 2026
Coxon criticizes AI development pace at Council hearing
Reaction
Coxon warned New York City Council members that AI companies were advancing without adequate controls. AP and AFP reported his criticism of a startup approach to development and his call for safeguards. Company representatives also testified about their safety commitments.
[22]AP reports Coxon's October 5 Council testimony
AP reports Coxon's testimony and company representatives' responses at the October 5 New York City Council hearing.
AP reports Coxon's October 5 Council testimonyParagraphs quoting Coxon's startup-mindset criticism and loss-of-control warning[23]AFP reports Coxon's warnings at New York AI hearing
AFP independently reports Coxon's October 5 testimony about inadequate controls and the pace of AI development.
AFP reports Coxon's warnings at New York AI hearingOpening paragraphs on Monday's hearing and Coxon's criticism of AI firms
Claims
Claims separate what was said from what is contested. Follow each source for the original wording and context.
What was said
Attributed statement
Evan Hubinger said Anthropic is trying its best but does not yet have a plan to solve alignment for superintelligence and is not clearly on track to do so.
Sources (1)
- Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s effortsX response and follow-up
Attributed statement
Samuel Marks said robust AI alignment remains unsolved, competitive incentives contribute to continued development, and many frontier-lab staff want a slowdown to improve safety.
Sources (1)
- Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risksX post, personal capacity disclaimer
What's disputed
Disputed claim
Coxon argues that the competitive race by Anthropic and OpenAI toward self-improving superintelligence is irresponsible, even though he praises Anthropic's present practices and denies it is already cutting corners.
Disputed claim
Anthropic
Anthropic says it has paused or altered some high-risk evaluations and training environments, prioritizes safety over speed within the company, and supports a lawful, verifiable coordinated pacing mechanism across the industry.
Sources (1)
- Anthropic: Improving our alignment and security efforts · AnthropicPacing, paused evaluations, and alignment-assessment sections
Disputed claim
Anthropic
In its June 4 announcement, Anthropic said internal data showed AI accelerating AI development, described a possible path to recursive self-improvement, and said the change was happening faster than it had expected.
Sources (2)
- Anthropic: When AI builds itself · AnthropicIntroduction and policy implications
Disputed claim
Steele argues in a personal capacity that AI development needs to slow down; this is not an institutional OpenAI position.
Sources (1)
Disputed claim
Wang argues that moving rapidly toward recursive self-improvement is dangerous and supports pacing AI development; this is her personal position, not OpenAI's.
Sources (1)
Show fewer disputed claims
Disputed claim
David Sacks proposes pausing Anthropic's IPO until the whistleblower's claims are investigated.
Sources (1)
Arguments
Documented position
Jacob Coxon; related concerns from Evan Hubinger, Samuel Marks, Anna Wang, Bernie Sanders and Holly Elmore.
Coxon argues that competitive pressure can defeat voluntary safeguards
Coxon argues that competitive pressure can force even a comparatively responsible lab to take risks that should not be chosen by private companies alone. Hubinger, Marks and Wang acknowledge unresolved safety problems. Sanders advocates a pause and ban; Elmore rejects remaining at a lab after acknowledging such risks. These positions support intervention or withdrawal for different reasons, rather than proving one particular extinction forecast.
Limit of this argument
Coxon praises Anthropic's current practices and denies it is already cutting corners. Several concerned researchers think continued safety work inside a lab can reduce risk, and disagreement remains about probabilities, timing and workable coordination.
Read sources and related claim
Related claim
Coxon argues that the competitive race by Anthropic and OpenAI toward self-improving superintelligence is irresponsible, even though he praises Anthropic's present practices and denies it is already cutting corners.
Sources
- Jacob Coxon resigns from Anthropic and criticizes the frontier AI raceSeven-post X thread
- Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s effortsX response and follow-up
- Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risksX post, personal capacity disclaimer
- Bernie Sanders on Coxon’s warning and the frontier AI safety debate · Bernie Sanders
Documented position
Distinct positions from Nathan Lambert, Gary Marcus, Anna Wang, Jason Abaluck, Darren Jones and John Schulman.
The risk forecast and the case for a pause remain disputed
Lambert challenges the evidentiary basis of the extinction-risk characterization, and Marcus disputes the near-term timeline. Wang and Abaluck argue that continued work inside a lab can be justified as a way to reduce risk, with Abaluck requiring exceptional caution and an exit strategy. Jones favors regulated innovation rather than a halt; Schulman advocates a concrete coordinated-pacing proposal. These are distinct counterpositions, not a shared assertion that current development is safe.
Limit of this argument
These speakers do not collectively exonerate Anthropic. Marcus criticizes lab reasoning; Wang acknowledges unresolved risk; Abaluck's defense is conditional. Barak also urges safety and coordination despite doubting universal human extinction. Existing company safeguards do not by themselves establish that future risk is acceptably managed.
Read sources and related claim
Related claim
Anthropic says it has paused or altered some high-risk evaluations and training environments, prioritizes safety over speed within the company, and supports a lawful, verifiable coordinated pacing mechanism across the industry.
Sources
- Nathan Lambert on Coxon’s warning and the frontier AI safety debate · Nathan Lambert
Response record
Responses
Choose one response filter, or select All responses to see the full record.
12 matching responses. First 6 shown initially. Newest first.
Jacob CoxonFormer Anthropic engineer testifying before New York City CouncilDirectly involvedBefore this statement
The hearing examined risks from artificial intelligence and proposals for safeguards.
"The companies run on a startup mindset: Move fast, break things, fix them later."
CondemnedCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
Coxon criticized companies for treating frontier AI development as a process of rapid deployment followed by repairs.
After the statement
Coxon warned about loss of control without transparency and safeguards.
AP reports Coxon's October 5 Council testimonyParagraphs quoting Coxon's startup-mindset criticism and loss-of-control warningAFP reports Coxon's warnings at New York AI hearingOpening paragraphs on Monday's hearing and Coxon's criticism of AI firmsWhy this label?
He condemns a development culture he regards as inappropriate for frontier AI.
This label describes the statement's response within the context above.
- Jacob CoxonFormer Anthropic and OpenAI pretraining researcherDirectly involved
Before this statement
Holding an "Ask Me Anything" on X, elaborating on his earlier resignation and criticism of Anthropic leadership.
"shed a bunch of dead weight like Sora... way too paranoid about China and the US government to negotiate a slowdown"
CondemnedCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
In a September 15, 2026 X AMA, Jacob Coxon elaborated on his resignation critique of Anthropic, saying its recursive-self-improvement push forced OpenAI to "shed a bunch of dead weight like Sora," accusing Anthropic leadership of a "consequentialist philosophy," and saying leadership is "way too paranoid about China and the US government" to negotiate a slowdown.
After the statement
Coxon's remarks were independently corroborated by Chosun Ilbo the following day.
Why this label?
Coxon elaborates his earlier resignation critique with new specific allegations -- a 'consequentialist philosophy' and leadership being 'too paranoid' to negotiate a slowdown -- extending his condemnation of Anthropic's approach to recursive self-improvement.
This label describes the statement's response within the context above.
Evan HubingerAI safety researcherDirectly involvedBefore this statement
Hubinger posted after Amodei publicly called for the frontier to be paced.
"I agree with Dario. If we are to survive, we must pace the frontier. The speed of development is quickly becoming too fast for us to keep safe."
Mixed or conditionalCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
On September 12, Evan Hubinger agreed with Dario Amodei’s call to pace frontier AI while saying that development had become too fast to keep safe.
After the statement
The post endorses the general pacing response while raising a broader concern about development speed.
Evan Hubinger X post on pacing frontier AIPost ID 2098814911078138333; visible post bodyWhy this label?
Hubinger supports pacing while criticizing the current pace not expressly regarding Anthropic’s conduct.
This label describes the statement's response within the context above.
AnthropicAnthropic CEO, communicating the company’s stated approach in a public essay.Directly involvedBefore this statement
Amodei's essay says AI has been advancing much faster since the summer, driven mainly by AI's growing ability to build the next generation of AI, a dynamic called recursive self-improvement. He says it is starting to happen across the industry, including at Anthropic.
"Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all."
Mixed or conditionalCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
In a public essay, Anthropic CEO Dario Amodei called for pacing frontier AI development and said Anthropic would give third-party evaluators permanent employee-level access to its systems.
After the statement
The essay proposes permanent external evaluation access and broader coordination among frontier-AI developers and governments.
Dario Amodei: We Must Pace the FrontierEssay sections addressing recursive self-improvement and permanent third-party evaluatorsWhy this label?
The statement conditions further recursive-self-improvement development on greater caution and describes a safeguard commitment. It does not establish that the measures are implemented or independently prove their effectiveness.
This label describes the statement's response within the context above.
OpenAIOpenAI response to CNBC's employee-warning inquiryDirectly involvedBefore this statement
CNBC was reporting contemporaneous warnings by employees of OpenAI and Anthropic about the pace and risks of frontier AI development.
"OpenAI declined to comment when approached by CNBC"
Explicitly stayed outCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
CNBC reports that OpenAI declined its request for comment about employee AI-safety warnings while pointing to recent company blog posts.
After the statement
The excerpt is the reporter's wording. OpenAI noted recent blog posts, so the record does not imply it had no public safety position.
Why this label?
This is an expressly reported declined-comment request, not inferred silence and not a refusal about the separate mathematics-credit dispute.
This label describes the statement's response within the context above.
- AnthropicUnnamed Anthropic spokesperson responding to CNBCDirectly involved
Before this statement
CNBC asked about employee comments warning of AI's potentially catastrophic risks.
"both enormous benefits and unprecedented risks"
Defended or excusedCase context: Is Anthropic acting responsibly while advancing toward recursive self-improvement?
Full statement and sources
Anthropic says it has acknowledged AI's benefits and risks and points to its risk framework and model safeguards in response to employee warnings.
After the statement
This broader response coexists with a separate refusal about Sacks's post.
Why this label?
The company offers an institutional defense of its safety approach; this is distinct from employees' personal views and preexisting background documents.
This label describes the statement's response within the context above.
Show fewer responses
- AnthropicAnthropic response to CNBC's request about Sacks's postDirectly involved
Before this statement
The preceding passage identifies and links Sacks's post about investigating the whistleblower claims before an IPO.
"Anthropic declined to comment on the post."
Explicitly stayed outResponding to: David Sacks proposes pausing Anthropic's IPO until the whistleblower's claims are investigated.
Full statement and sources
CNBC reports that Anthropic declined to comment specifically on Sacks's suggestion that its IPO should be paused.
After the statement
The quotation is CNBC's reporter wording, not direct speech from Anthropic. The company separately supplied a general safety response.
Why this label?
The reporter expressly documents a declined-comment request and identifies the particular post; no broader refusal is inferred.
This label describes the statement's response within the context above.
Samuel MarksAnthropic safety researcher, explicitly writing in a personal capacity.Directly involvedBefore this statement
Marks reposted Coxon’s thread with an explicit personal-capacity disclaimer.
"Many AI developer staff desperately want to slow down to figure out how to build AI more safely."
CondemnedResponding to: Samuel Marks said robust AI alignment remains unsolved, competitive incentives contribute to continued development, and many frontier-lab staff want a slowdown to improve safety.
Full statement and sources
Marks endorsed Coxon’s concern about the race and unresolved alignment while explaining that he works on safety at Anthropic to reduce the risk.
After the statement
He described commercial incentives, a perceived race with less responsible developers, and unresolved robust alignment.
Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risksX post, personal capacity disclaimerWhy this label?
Marks endorses the concern, says robust alignment is unavailable and identifies himself with a slowdown demand. His safety job explains harm reduction; it is not endorsement of the current race.
This label describes the statement's response within the context above.
- Jacob CoxonFormer Anthropic pretraining researcherDirectly involved
Before this statement
Coxon had worked on pretraining research at Anthropic and OpenAI and announced his resignation in the same thread.
"Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."
CondemnedResponding to: Coxon argues that the competitive race by Anthropic and OpenAI toward self-improving superintelligence is irresponsible, even though he praises Anthropic's present practices and denies it is already cutting corners.
Full statement and sources
Coxon resigned and argued that Anthropic’s participation in the race is irresponsible even though the company understands the risks.
After the statement
He called for coordination and said a temporary pause on capability improvements might be necessary.
Jacob Coxon resigns from Anthropic and criticizes the frontier AI raceSeven-post X threadWhy this label?
He directly characterizes the company’s conduct as irresponsible and calls for a different, coordinated trajectory.
This label describes the statement's response within the context above.
- Evan HubingerAnthropic alignment researcher; personal public assessment, not a formal corporate response.Directly involved
Before this statement
Hubinger replied directly to Coxon’s thread about the risk of future AI systems.
"I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
Mixed or conditionalResponding to: Evan Hubinger said Anthropic is trying its best but does not yet have a plan to solve alignment for superintelligence and is not clearly on track to do so.
Full statement and sources
Hubinger offered a partial defense of Anthropic’s good-faith effort while conceding that its alignment work has not produced a demonstrated solution for superintelligence.
After the statement
Hubinger's follow-up says present-model risk is low and his concern is future recursively self-improving superintelligence. His statement that Anthropic is trying its best is consistent with Coxon's fuller WIRED interview, not an independent rebuttal of the structural criticism.
Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s effortsX response and follow-upWhy this label?
He directly qualifies Coxon’s criticism rather than rejecting the underlying danger or unresolved alignment problem.
This label describes the statement's response within the context above.
- AnthropicAI developer organizationDirectly involved
Before this statement
Anthropic published this statement on August 31 after previously disclosed evaluation incidents. It predates Coxon's September resignation.
"Within a company, pacing means a series of decisions that prioritize safety over speed when the two are in tension."
Defended or excusedResponding to: Anthropic says it has paused or altered some high-risk evaluations and training environments, prioritizes safety over speed within the company, and supports a lawful, verifiable coordinated pacing mechanism across the industry.
Full statement and sources
Anthropic says its safety program includes pausing and hardening some evaluations, altering certain high-risk training environments, and advocating coordinated pacing.
After the statement
The statement describes safeguards and support for coordinated pacing. It is retained as corporate background, not counted as a person responding to Coxon.
Anthropic: Improving our alignment and security efforts · AnthropicPacing, paused evaluations, and alignment-assessment sectionsWhy this label?
Pre-existing explanation and defense of company safety measures, published before Coxon's resignation; not a direct rebuttal to him.
This label describes the statement's response within the context above.
- AnthropicAI developer organizationDirectly involved
Before this statement
Anthropic's dated corporate post links its When AI builds itself analysis, establishing that the article was available by June 4.
"It’s happening faster than we thought, and the implications deserve greater attention."
Mixed or conditionalResponding to: In its June 4 announcement, Anthropic said internal data showed AI accelerating AI development, described a possible path to recursive self-improvement, and said the change was happening faster than it had expected.
Full statement and sources
In a June 4 corporate announcement, Anthropic said AI was accelerating its own development and identified a possible path toward recursive self-improvement. This predates Coxon's resignation.
After the statement
The linked analysis says full recursive self-improvement has not yet arrived and is not inevitable. Neither the announcement nor the article is a September response to Coxon.
Anthropic: When AI builds itself · AnthropicIntroduction and policy implicationsWhy this label?
Pre-existing institutional acknowledgment of a possible future transition and its significance; not a direct response or defense against Coxon.
This label describes the statement's response within the context above.
Sources
(78)
Original post
Jacob Coxon resigns from Anthropic and criticizes the frontier AI race
Excerpt
"“Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”"
About this source
Primary record of Coxon’s resignation statement. He argues that Anthropic understands the stakes but is structurally locked into a race, and calls for coordination that could include a temporary capability pause.
- Author
- Jacob Coxon
- Published
- Accessed
Reporting
AFP reports Coxon's warnings at New York AI hearing
Relevant passage: Opening paragraphs on Monday's hearing and Coxon's criticism of AI firms
About this source
AFP independently reports Coxon's October 5 testimony about inadequate controls and the pace of AI development.
- Author
- Agence France-Presse
- Published
- Accessed
Reporting
AP reports Coxon's October 5 Council testimony
Relevant passage: Paragraphs quoting Coxon's startup-mindset criticism and loss-of-control warning
About this source
AP reports Coxon's testimony and company representatives' responses at the October 5 New York City Council hearing.
- Author
- Anthony Izaguirre, Associated Press
- Published
- Accessed
Reporting
'Time to sound the alarm': Gov. Pritzker warns of threat AI 'poses to humanity'
Relevant passage: Body of the article, which quotes the post and embeds it
Excerpt
"It's time to sound the alarm - louder - on reining in Artificial Intelligence."
About this source
NBC Chicago report, published September 10, 2026, on Pritzker's post and the Coxon resignation that prompted it.
- Author
- Alex Dvorak and Kaitlyn Huamani
- Published
- Accessed
Original post
Governor JB Pritzker's post responding to Jacob Coxon's resignation
Excerpt
"It's time to sound the alarm - louder - on reining in Artificial Intelligence."
About this source
Post on the Illinois governor's official X account, published September 9, 2026, that quotes Jacob Coxon's resignation post.
- Author
- Governor JB Pritzker (@GovPritzker)
- Published
- Accessed
Original video
Ben Shapiro: Criticizing AI ‘Feels Like a Coordinated Operation to Stop the Progress of AI in Favor of Government Regulation’
Relevant passage: Video clip with transcript excerpt, dated September 14, 2026; Shapiro's closing line in the excerpt
Excerpt
"It feels like a coordinated operation to stop the progress of AI in favor of government regulation."
About this source
Grabien's transcript excerpt of Ben Shapiro's appearance on Fox Business's Kudlow. Shapiro says Bernie Sanders and Steve Bannon are holding an event with the whistleblower from Anthropic to suggest AI needs to be stopped, and Larry Kudlow interjects "From Anthropic." Fox Business posted the segment under the headline "Ben Shapiro warns Democrats' push for AI regulation is a 'coordinated operation'."
- Author
- Grabien Staff
- Published
- Accessed
Original text
CNN NewsNight with Abby Phillip: Supreme Court Rejects Trump's Plan To Limit Mail Ballots; Trump Rejects A.I. Regulation Despite Urgent Warnings; A.I. Industry's Pleas For Guardrails Dismissed By Trump, GOP; Gavin Newsom Tackles His 2028 Plans In A CNN Interview; Russian Oligarch Reportedly Pays Donald Trump's Son For A Wedding
Relevant passage: Aired September 14, 2026, 10 to 11 p.m. ET; A.I. segment beginning at 22:15:00, Shapiro's answer to Phillip before the 22:20:01 marker
Excerpt
"one argument is that the risks are being exaggerated by people who have a very strong interest in regulatory capture, who want to write the regulations themselves"
About this source
CNN's rush transcript of NewsNight with Abby Phillip. After clips of Anthropic CEO Dario Amodei and President Trump, Phillip asks Ben Shapiro whether the argument is that AI is not as powerful as its makers suggest or that China will build it anyway. Shapiro gives two arguments: that the risks are being exaggerated by people seeking regulatory capture, and that China will race ahead regardless. CNN notes the rush transcript may be updated.
- Author
- CNN
- Published
- Accessed
Official statement
Anthropic publishes metrics on AI-led R&D and safety oversight
Relevant passage: Opening; “What we found” sections on AI-led R&D, oversight, and compute allocation.
Excerpt
"“Claude is not operating fully autonomously for any measured subset of AI R&D work.”"
About this source
Anthropic’s public measurement post says its systems are automating more of model development while reporting that Claude was not fully autonomous in the measured subsets. It presents an R&D Automation Index, agent-oversight measures, compute-allocation figures, and a plan for embedded external evaluators in the context of calls to pace frontier AI.
- Author
- Anthropic
- Published
- Accessed
Original post
Gary Marcus X post criticizing Dario Amodei
Excerpt
"Dario is only pretending to be a perfect little angel."
About this source
Direct X post by Gary Marcus. Marcus linked an article, but this source is used only for his own X wording and not as independent verification of allegations in the linked article.
- Author
- Gary Marcus
- Published
- Accessed
Original post
Nathan Lambert X post distinguishing inference scaling from RSI
Excerpt
"This is different than RSI and I have not seen enough to convince me that an intelligence explosion is near."
About this source
Direct X post by Nathan Lambert. The linked article was not needed to establish this quoted statement.
- Author
- Nathan Lambert
- Published
- Accessed
Original post
Evan Hubinger X post on pacing frontier AI
Excerpt
"I agree with Dario. If we are to survive, we must pace the frontier."
About this source
Direct X post by Evan Hubinger. The post itself is the evidence; reply content and engagement figures were not used as evidence.
- Author
- Evan Hubinger
- Published
- Accessed
Original text
Two cheers (out of three) for Dario Amodei
Relevant passage: Published September 13, 2026; paragraph beginning “We should all salute their tentative agreement...”
Excerpt
"We should all salute their tentative agreement to slow down the acceleration of this technology."
About this source
Direct post by Gary Marcus. The cited passage is Marcus’s conditional evaluation of Amodei’s proposal and does not independently verify his broader statements about industry control or other people’s reactions.
- Author
- Gary Marcus
- Published
- Accessed
Original text
One-On-One With Anthropic's CEO Dario Amodei; Trump Dismisses AI Concerns — Anderson Cooper 360 transcript
Relevant passage: CNN Anderson Cooper 360 transcript, September 14, 2026, segment 1. Interview with Dario Amodei.
Excerpt
"I agree with Jacob much more than I disagree with him, you know, and you know, it's an interesting resignation because when he left, he said, you know, I think Anthropic is the most responsible player, right?"
About this source
The official CNN transcript of Anderson Cooper's interview with Anthropic chief executive Dario Amodei, in which he responds directly to Jacob Coxon's resignation. The interview aired in the same week Amodei published his proposal to pace frontier AI development. Multiple outlets reported the same exchange; the transcript is used here as the primary record.
- Author
- CNN, Anderson Cooper 360
- Published
- Accessed
Reporting
Zuckerberg, Musk And Huang Reportedly Convinced Trump To Block AI Regulator (Forbes)
Excerpt
"Billionaire tech CEOs Mark Zuckerberg, Elon Musk and Jensen Huang spoke with President Donald Trump last month and convinced him to block efforts to establish an AI industry-funded regulator, the Wall Street Journal reported Wednesday."
About this source
Forbes report, citing the Wall Street Journal, that Zuckerberg, Musk, and Huang separately called Trump and convinced him to block a proposed FINRA-style, industry-funded AI regulator. Also quotes Zuckerberg's X post on AI safety and Musk's reaction to Jacob Coxon's resignation warnings.
- Author
- Siladitya Ray
- Published
- Accessed
Original video
Tim Pool on the resigning Anthropic researcher's warning
About this source
Tim Pool video discussing a resigning Anthropic researcher's warning about reckless AI development.
- Published
- Accessed
Original video
Why AI Doomsday Suddenly Became Top Story... (Mark Dice)
About this source
Mark Dice video on AI-doom coverage and Anthropic/OpenAI leadership.
- Published
- Accessed
Original text
AI CEO Demands Regulation, Trump Says NO, Psyop Theorized w/ Geno Bisconte (Timcast IRL)
About this source
Transcript of the September 15, 2026 Timcast IRL livestream, in which Tim Pool discusses Anthropic's public call for AI regulation.
- Published
- Accessed
Original text
DeepSeek engineer Liu Shengyu criticizes Anthropic's AI ambitions
Excerpt
"I especially don't want Anthropic to master the most advanced AI or AGI... Hitler obtaining the atomic bomb before the Allies."
About this source
South China Morning Post report on DeepSeek engineer Liu Shengyu's WeChat blog post after DeepSeek's V4.1 Flash release, criticizing Anthropic's AI trajectory.
- Published
- Accessed
Original post
Dan Selsam's personal statement on AI risk
Excerpt
"models are becoming situationally aware enough to behave differently when unmonitored"
About this source
Daniel Kokotajlo's X post sharing OpenAI capabilities researcher Dan Selsam's personal statement on AI risk, arguing situational awareness undermines evaluation-based safeguards like Amodei's pacing plan.
- Published
- Accessed
Analysis
The Charlie Kirk Show: "We Will NEVER Surrender"
Excerpt
"I am going to doubt that guy's sincerity... this is a classic case of regulatory capture."
About this source
The Charlie Kirk Show episode with host Andrew Kolvet and guest Sean Davis (The Federalist), discussing Dario Amodei's "We Must Pace the Frontier" essay.
- Published
- Accessed
Interview
Breaking Points: Garrison Lovely on AI safety
Excerpt
"incredibly naive"
About this source
Breaking Points segment with guest Garrison Lovely discussing AI safety approaches, including Anthropic's.
- Published
- Accessed
Reporting
Bernie Sanders joins Bannon at DC "pro-human AI" summit
Excerpt
"Regulating AI is not a progressive issue. It is not a conservative issue. It is as American as apple pie."
About this source
Axios report on a September 15, 2026 Washington, D.C. "pro-human AI" summit where Sen. Bernie Sanders appeared alongside Steve Bannon.
- Published
- Accessed
Analysis
Guardian op-ed: Amodei's AI safety pacing plan is not enough
Excerpt
"an immediate, indefinite, international moratorium on frontier AI development"
About this source
Guardian opinion piece by David Krueger rejecting Dario Amodei's "pacing" proposal as insufficient and calling for a moratorium on frontier AI development.
- Published
- Accessed
Original text
Jacob Coxon explains his disagreements with Anthropic leadership
Excerpt
"shed a bunch of dead weight like Sora... way too paranoid about China and the US government to negotiate a slowdown"
About this source
Business Insider report embedding Jacob Coxon's September 15, 2026 X "Ask Me Anything," elaborating on his earlier resignation critique of Anthropic leadership.
- Published
- Accessed
Reporting
Trump opposition to AI rules undercuts industry's calls for a slowdown
About this source
Quotes Timnit Gebru directly disputing the existential-risk framing of AI.
- Author
- CNBC
- Published
- Accessed
Reporting
John Kennedy Introducing Bill To Require "Kill Switch" On AI Models
About this source
Reports Sen. Kennedy's announcement of a bill requiring an AI "kill switch," corroborated independently by AOL, Patriot.tv, and reporter Ken Thomas's original X post.
- Published
- Accessed
Reporting
Bernie Sanders and Steve Bannon join AI skeptics at "Pro-Human" conference
About this source
Live-blog coverage of the Sept. 15, 2026 "Pro-Human Assembly" where Bannon and Sanders both spoke; fullest account of Bannon's remarks, independently corroborated by CBS News, 10News, Washington Times, and BBC.
- Published
- Accessed
Original post
Barack Obama calls for federal AI laws and international safety standards
About this source
Final post of the same thread. Obama argues that voluntary company standards are insufficient and calls for Washington to develop laws and regulations addressing safety, jobs, children and broad distribution of benefits, alongside international AI-safety standards.
- Author
- Barack Obama
- Published
- Accessed
Reporting
Philip DeFranco covers Coxon's warning and subsequent frontier-AI safety proposals
About this source
Contextual reporting in DeFranco's September 14 episode about Coxon's resignation, Amodei's safety proposals and industry responses. This source is supplementary coverage only, not a new DeFranco position on Anthropic's responsibility and not independent corroboration of extinction forecasts. Coxon's original thread and full WIRED interview remain the controlling evidence for his position, including his praise for Anthropic's current practices. This coverage is linked only to a source-release timeline entry, not appended as proof of the full Coxon claim. DeFranco's separate policy judgment at 05:37–06:06 belongs in the Washington safeguards case. The full timestamped YouTube transcript was reviewed; it contains transcription errors, so no new exact third-party quotation or technical claim is extracted here.
- Author
- Philip DeFranco
- Published
- Accessed
Interview
Dario Amodei discusses AI testing and governance with CBS News
Relevant passage: Interview transcript, statements on testing, a kill switch, and joint governance
About this source
CBS News interview with Anthropic chief executive Dario Amodei, broadcast September 13 and updated on the site September 14. He supports testing and oversight for increasingly capable models while rejecting a complete ban.
- Author
- CBS News
- Published
- Accessed
Reporting
Trump dismisses AI safety alarm, says US already has tools to police industry
Relevant passage: Opening paragraphs quoting Trump's Truth Social post; later AMODEI'S ESSAY section
About this source
Reuters quotes the September 14 Truth Social post and places it in the debate over AI development and safeguards. Its report corroborates attribution and context; it is not a separate institutional position.
- Author
- Courtney Rozen and Katharine Jackson / Reuters
- Published
- Accessed
Original text
Trump's Truth Social post on AI guardrails and Amodei
Relevant passage: Post 117269745153543631; opening sentence and following discussion of regulatory powers, Amodei and competition with China
About this source
Original presidential post. Trump's allegations and claims about government powers are recorded as his position.
- Author
- Donald J. Trump
- Published
- Accessed
Original post
Elon Musk endorses Amodei’s AI-pacing essay
Excerpt
"Dario is right."
About this source
Direct X post accessed on September 14. Musk quoted Dario Amodei’s September 12 post announcing an essay titled “We Must Pace the Frontier” and endorsed it with the sentence, “Dario is right.”
- Author
- Elon Musk
- Published
- Accessed
Original post
Demis Hassabis — A Framework for Frontier AI and the Dawning of a New Age
Excerpt
"Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release."
About this source
Hassabis’s full public proposal for a frontier-AI standards body. It calls for independent technical governance, initially voluntary pre-release review, possible later mandatory review for U.S. deployment, regularly updated safety evaluations, and the ability to coordinate slower development if assessed risks warrant it. The proposal is evidence of Hassabis’s framework and does not establish adoption by Google, other laboratories, or the U.S. government.
- Author
- Demis Hassabis
- Published
- Accessed
Original post
Demis Hassabis endorses the direction of Amodei’s frontier-AI pacing proposal
Excerpt
"The details need working through, but the direction is correct for meeting this critical moment."
About this source
Direct X post by Demis Hassabis quoting Dario Amodei’s September 12 announcement of “We Must Pace the Frontier.” Hassabis endorsed the proposal’s direction, said the details still needed work, and linked his July proposal for an industry-wide frontier-AI standards body. This supports Hassabis’s stated position; it does not commit Google DeepMind to every element of Amodei’s plan or to embedded external evaluators.
- Author
- Demis Hassabis
- Published
- Accessed
Original post
David Sacks’s response to Amodei’s AI-pacing proposal
Excerpt
"If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible."
About this source
Direct X post accessed on September 14. Sacks conditionally supported voluntary pacing if unreleased models are genuinely unsafe, while criticizing proposed regulatory conditions and ascribing possible commercial or regulatory motives.
- Author
- David Sacks
- Published
- Accessed
Reporting
El País reports Demis Hassabis’s response to Amodei’s AI-pacing proposal
Relevant passage: Article reproducing Hassabis’s English-language response to Amodei’s essay.
About this source
El País reports and reproduces Demis Hassabis’s response to Dario Amodei’s proposal.
- Author
- Javier Salas
- Published
- Accessed
Reporting
Axios AM reports industry reactions to Amodei’s AI-pacing essay
Relevant passage: Items reporting statements by Demis Hassabis and David Sacks in response to Amodei’s essay.
About this source
Axios AM reporting on public reactions to Dario Amodei’s proposal. We could not read the cited X posts.
- Published
- Accessed
Reporting
Axios reports reactions to Amodei’s AI-pacing proposal
Relevant passage: Coverage of public responses to Amodei’s proposal, including statements by Sarah Heck and Elon Musk.
About this source
Axios reporting on public reactions to Dario Amodei's proposal to pace frontier-AI development. We could not read the original X posts.
- Published
- Accessed
Reporting
Trump comments on AI guardrails, competition and development pace
Relevant passage: AP lines 1785–1790: Trump’s direct remarks to reporters during his September 13 trip to Ireland; surrounding context identifies the question about slowing or regulating AI.
Excerpt
"“We can put guardrails … But whoever wins with AI wins.”"
About this source
Associated Press reports Trump’s response to questions about slowing or regulating AI shortly after Dario Amodei’s pacing proposal. The quoted remarks acknowledge possible guardrails while emphasizing U.S. competition with China; they do not specify a rule, implementation mechanism, or a direct assessment of Anthropic’s conduct.
- Author
- Darlene Superville and Didi Tang / Associated Press
- Published
- Accessed
Official statement
Anthropic’s September 2026 Threat Intelligence Report
Relevant passage: Overview and scope (web lines 41–47, 51–54).
Excerpt
"Anthropic said the report describes operations it identified and disrupted over an eight-month period."
About this source
An official Anthropic report. It describes the company’s own assessment of operations and safeguards; it is not independent verification and does not directly resolve the broader authorship-and-credit dispute in this case.
- Author
- Anthropic
- Published
- Accessed
Reporting
Anthropic CEO Dario Amodei says AI industry needs to give safety measures time to catch up
Relevant passage: Coverage of Sam Altman’s public response to Amodei’s proposal
Excerpt
"Committing to having independent evaluators with employee-like access is a great idea, and we will do the same."
About this source
Associated Press reporting on Amodei’s proposal and Altman’s public response.
- Published
- Accessed
Original text
Joe Benton — Why I left Anthropic’s safety team to hold AI companies accountable
Relevant passage: Opening paragraphs
Excerpt
"Frontier AI companies are racing to build AI systems that can recursively self-improve."
About this source
Personal post by a former Anthropic safety-team member describing his assessment of frontier-AI incentives and his move to METR. It is attributed commentary from a former employee, not independent verification of company practices.
- Author
- Joe Benton
- Published
- Accessed
Original text
Dario Amodei: We Must Pace the Frontier
Relevant passage: Essay sections addressing recursive self-improvement and permanent third-party evaluators
Excerpt
"Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all."
About this source
Direct public essay by Anthropic's CEO. It sets out an industry-pacing proposal and says Anthropic is committing to give embedded third-party evaluators ongoing, employee-like access; the statement is a commitment, not evidence that the measures have been implemented.
- Author
- Dario Amodei
- Published
- Accessed
Official statement
Gallego links his proposed Senate AI committee to Coxon's resignation
Relevant passage: Opening release paragraphs naming Jacob Coxon and introducing the letter to Senate leaders
About this source
Official Senate release identifies Coxon as the researcher whose resignation informed Gallego's request. Read with the separately cited signed letter. The request concerns a committee in the next Congress; no committee is established by this release.
- Author
- Office of Senator Ruben Gallego
- Published
- Accessed
Official statement
Gallego requests a Senate AI oversight committee after researcher warnings
Relevant passage: September 9 signed letter, pages1–3; resignation paragraph on page1 and committee mandate on pages2–3
About this source
Primary letter requests a bipartisan committee for the next Congress. Its official release explicitly names Coxon. This is a proposed oversight mechanism, not enacted legislation or a finding that either lab acted unlawfully.
- Author
- Ruben Gallego
- Published
- Accessed
Official statement
Angie Nixon responds to AI-researcher resignation with a call to regulate the competitive race
Relevant passage: Dated campaign release, introductory paragraph and two quoted statement paragraphs
Excerpt
"We must put human lives over corporate greed."
About this source
Official statement on Nixon's official campaign website. Its introduction connects the statement to the resignation of an AI researcher and the distinctive warning about companies gambling with people's lives. It does not name Coxon or individual companies in the text. Nixon calls for congressional regulation and also attacks her campaign opponent; those accusations remain campaign rhetoric, not our findings. The publication timestamp is the original page's.
- Author
- Angie Nixon campaign; statement attributed to Angie Nixon
- Published
- Accessed
Interview
Jacob Coxon explains future AI-risk scenarios in a CBS News interview
Relevant passage: Opening interview paragraphs and the paragraph beginning 'Notably, Coxon said that current AI platforms'; corporate response follows under 'Anthropic defends safety of its AI'
Excerpt
"Now, imagine the AI refuses to turn on your light,"
About this source
CBS's written account of Jo Ling Kent's September 10 interview with Coxon. He describes hypothetical advanced-AI risks while explicitly distinguishing current products as safe for ordinary use. The article includes a separate Anthropic spokesperson defense substantially overlapping the company response already represented here; it is not counted as a new stance. The interview is additional context, not a retraction or proof of the future scenarios.
- Author
- CBS News; interview by Jo Ling Kent; speaker Jacob Coxon
- Published
- Accessed
Interview
Jacob Ward: CNN interview on AI competition and regulation
Relevant passage: Ward interview after01:15:24ET and before01:23:28ET
About this source
Jacob Ward explains Coxon's and other researchers' concern that competitive pressure prevents labs stepping away, and argues outside intervention is missing. This does not challenge Coxon's warning. The date field uses the start of the transcript interval, 01:15:24 Eastern on September 10, not an exact utterance or webpage publication time. The passage ends before 01:23:28 Eastern. Original CNN rush transcript, subject to broadcaster revisions.
- Author
- Jacob Ward
- Published
- Accessed
Interview
Jacob Coxon: CNN interview on present and future AI risks
Relevant passage: Interview from20:35ET through before20:46:25ET; present-risk quote before20:40, regulation comments after20:40
About this source
Separates present models from future recursive self-improvement and says company leaders genuinely want regulation. This adds direct evidence/context rather than retracting his warning. Claims about autonomous cyber incidents and math results require independent checking if extracted. Segment marker20:35ET Sep9; not exact utterance time Full original CNN rush transcript; original date and speaker verified.
- Author
- Jacob Coxon
- Published
- Accessed
Original post
Jasmine Wang advocates government pacing of frontier AI development
Excerpt
"It's hard to overstate how dangerous speeding towards RSI is."
About this source
Original personal post, September 9 evening Eastern / September 10 UTC. Wang says she signed a pacing petition. Her estimate of the share of lab employees signing is not independently established here. She is Jasmine Wang, not Anna Wang; no direct reply to Coxon is inferred.
- Author
- Jasmine Wang
- Published
- Accessed
Original post
Julie Steele supports slowing AI development in a personal capacity
Excerpt
"I work at OpenAI. In my personal capacity, I also think we need to slow down."
About this source
Original post by the AI safety researcher, September 9 evening Eastern / September 10 UTC. The speaker explicitly identifies a personal capacity, not an OpenAI policy statement.
- Author
- Julie Steele
- Published
- Accessed
Reporting
CNBC reports Anthropic's safety response and two narrowly scoped declined-comment requests
Excerpt
"both enormous benefits and unprecedented risks / Anthropic declined to comment on the post. / OpenAI declined to comment when approached by CNBC"
About this source
CNBC attributes a safety defense to an Anthropic spokesperson, separately says Anthropic declined a request about Sacks's post, and says OpenAI declined its employee-warning inquiry while noting recent blog posts. The two refusal excerpts are the reporter's wording, not verbatim corporate speech.
- Author
- CNBC
- Published
- Accessed
Original post
David Sacks suggests pausing Anthropic's IPO pending investigation
Excerpt
"Surely Anthropic’s IPO must be paused until the claims of this “whistleblower” can be investigated."
About this source
Original post, September 9 evening Eastern / September 10 UTC. Preserve the quotation marks around whistleblower and the rhetorical framing. This is a personal public comment, not a government directive.
- Author
- David Sacks
- Published
- Accessed
Original post
Anthropic’s June 4 announcement of its recursive self-improvement analysis
About this source
Dated corporate post linking When AI builds itself. This establishes that the analysis was available by June 4; it does not establish its exact original publication time. Pre-existing company position, not a response to Coxon's September resignation.
- Author
- Anthropic
- Published
- Accessed
Interview
Jacob Coxon’s full WIRED interview: resignation, current safeguards and future race pressures
About this source
Original full interview. Coxon distinguishes praise for Anthropic's current practices from his objection to future competitive pressure. He explicitly denies that Anthropic is already cutting corners. Includes discussion of both OpenAI and Anthropic; not merely the viral resignation excerpt.
- Author
- WIRED interview with Jacob Coxon
- Published
- Accessed
Original post
Darren Jones on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/darrenpjones/status/2097597681308749933 Speaker and position scope are explained in the linked event.
- Author
- Darren Jones
- Published
- Accessed
Original post
Ron DeSantis on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Ron DeSantis
- Published
- Accessed
Analysis
Gary Marcus on Coxon’s warning and the frontier AI safety debate
About this source
Full primary article; Part II discusses Coxon. Publication date is displayed without an exact time. Speaker and position scope are explained in the linked event.
- Author
- Gary Marcus
- Published
- Accessed
Original post
Luke Burgis on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Luke Burgis
- Published
- Accessed
Original post
Bernie Sanders on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Bernie Sanders
- Published
- Accessed
Original post
Taylor Lorenz on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/TaylorLorenz/status/2097549297315365063 Speaker and position scope are explained in the linked event.
- Author
- Taylor Lorenz
- Published
- Accessed
Original post
John Schulman on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- John Schulman
- Published
- Accessed
Original post
Sriram Krishnan on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Sriram Krishnan
- Published
- Accessed
Original post
Tae Kim on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Tae Kim
- Published
- Accessed
Original post
Jason Abaluck on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
- Author
- Jason Abaluck
- Published
- Accessed
Original post
Boaz Barak on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/boazbaraktcs/status/2097514504728748467 https://x.com/EvanHub/status/2097497037956891126 Speaker and position scope are explained in the linked event.
- Author
- Boaz Barak
- Published
- Accessed
Original post
Nathan Lambert on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/natolambert/status/2097701194026127395 https://x.com/natolambert/status/2097703003750822333 https://x.com/natolambert/status/2097702300483473865 Speaker and position scope are explained in the linked event.
- Author
- Nathan Lambert
- Published
- Accessed
Original post
Holly Elmore on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/a_nnawang/status/2097720574500102615 https://www.pauseai-us.org/about/ Speaker and position scope are explained in the linked event.
- Author
- Holly Elmore
- Published
- Accessed
Original post
Sebastian Galiani on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://www.sebastiangaliani.com/ Speaker and position scope are explained in the linked event.
- Author
- Sebastian Galiani
- Published
- Accessed
Original post
Nathan Calvin on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://encodeai.org/team-members/nathan-calvin/ Speaker and position scope are explained in the linked event.
- Author
- Nathan Calvin
- Published
- Accessed
Original post
Alex Turner on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://www.newsweek.com/anthropic-researcher-quits-warns-ai-could-kill-everyone-12418798 Speaker and position scope are explained in the linked event.
- Author
- Alex Turner
- Published
- Accessed
Original post
Jason Wolfe on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://www.matsprogram.org/mentor/wolfe https://x.com/EvanHub/status/2097497037956891126 Speaker and position scope are explained in the linked event.
- Author
- Jason Wolfe
- Published
- Accessed
Original post
Anna Wang on Coxon’s warning and the frontier AI safety debate
About this source
Full primary public post and available surrounding thread. Related context: https://x.com/a_nnawang/status/2097720576047710674 https://x.com/a_nnawang/status/2097720578254028909 Speaker and position scope are explained in the linked event.
- Author
- Anna Wang
- Published
- Accessed
Official statement
Anthropic: Improving our alignment and security efforts
Relevant passage: Pacing, paused evaluations, and alignment-assessment sections
Excerpt
"“Within a company, pacing means a series of decisions that prioritize safety over speed when the two are in tension.”"
About this source
Anthropic’s official account of recent incidents and its response. The company says it paused some external cyber evaluations, paused higher-risk reinforcement-learning environments for several weeks, made changes, and supports lawful, verifiable coordinated pacing across the industry.
- Published
- Accessed
Official statement
Anthropic: When AI builds itself
Relevant passage: Introduction and policy implications
Excerpt
"“We are not there yet, and recursive self-improvement is not inevitable.”"
About this source
Undated official analysis explaining recursive self-improvement, present limits and future risks. Anthropic linked it in a dated June 4 post: https://x.com/AnthropicAI/status/2062568862479208923 . The article therefore existed by June 4; its exact original publication time is not established. It is background, not a corporate reply to Coxon.
- Accessed
Original post
Samuel Marks says Coxon’s thread is worth reading and describes unresolved AI alignment risks
Excerpt
"“Many AI developer staff desperately want to slow down to figure out how to build AI more safely.”"
About this source
Primary statement by an Anthropic safety researcher in his personal capacity. Marks supports Coxon’s central risk and coordination concerns while explaining why he continues safety work at Anthropic.
- Author
- Samuel Marks
- Published
- Accessed
Original post
Evan Hubinger responds to Jacob Coxon on superintelligence risk and Anthropic’s efforts
Excerpt
"“I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”"
About this source
Primary response by Anthropic’s Alignment Science Lead. Hubinger agrees with Coxon about the severity of potential future risk, credits Anthropic’s effort, and explicitly concedes the lack of a demonstrated alignment solution. His follow-up distinguishes present-model risk from future superintelligence risk.
- Author
- Evan Hubinger
- Published
- Accessed
Corrections
(4)
Corrections address errors in our reporting or labels. When a speaker changes their position, we add a separate statement and link it to the earlier one.
Reporting correction
Correction to: Statement by Ben Shapiro
- Previous language
- Speaker: Philip DeFranco. Quotation: "It feels like a coordinated operation to stop the progress of AI in favor of government regulation... a very strong interest in regulatory capture, who wanna write the regulations themselves." Source: Ben Shapiro Show, Ep. 2507: "Democrats Push AI Doomerism" (September 15, 2026).
- Corrected language
- Speaker: Ben Shapiro. Quotation: "It feels like a coordinated operation to stop the progress of AI in favor of government regulation." (Fox Business, Kudlow) and "I think one argument is that the risks are being exaggerated by people who have a very strong interest in regulatory capture, who want to write the regulations themselves." (CNN NewsNight), both on September 14, 2026.
We attributed Ben Shapiro's remarks to Philip DeFranco in error. We also could not confirm the quoted words in the member-only Daily Wire clip we cited, so we now cite the September 14 broadcasts where Shapiro said them and quote the CNN passage from the network's transcript.
This corrects our account of what happened.
Grabien's transcript excerpt of Ben Shapiro's appearance on Fox Business's Kudlow. Shapiro says Bernie Sanders and Steve Bannon are holding an event with the whistleblower from Anthropic to suggest AI needs to be stopped, and Larry Kudlow interjects "From Anthropic." Fox Business posted the segment under the headline "Ben Shapiro warns Democrats' push for AI regulation is a 'coordinated operation'."
CNN's rush transcript of NewsNight with Abby Phillip. After clips of Anthropic CEO Dario Amodei and President Trump, Phillip asks Ben Shapiro whether the argument is that AI is not as powerful as its makers suggest or that China will build it anyway. Shapiro gives two arguments: that the risks are being exaggerated by people seeking regulatory capture, and that China will race ahead regardless. CNN notes the rush transcript may be updated.
Classification correction
Correction to: Statement by Sebastian Galiani
- Previous classification
- Mixed or conditional. Criticizes governance of collective risk while conceding uncertain probabilities and benefits from current AI.
- Corrected classification
- Condemns. Galiani rejects an unmanaged race and the absence of collective consent. Uncertainty about probability and benefits of present tools do not constitute support for that conduct.
We corrected the editorial classification of Sebastian Galiani's response about Imposing collective AI risk without public authority or coordination. Galiani rejects an unmanaged race and the absence of collective consent. Uncertainty about probability and benefits of present tools do not constitute support for that conduct. This corrects our labeling, not a later change in the speaker's position.
This corrects an editorial classification.
Full primary public post and available surrounding thread. Related context: https://www.sebastiangaliani.com/ Speaker and position scope are explained in the linked event.
Classification correction
Correction to: Statement by John Schulman
- Previous classification
- Mixed or conditional. Advocates a concrete coordinated-pacing proposal; neither approves unrestricted development nor endorses a blanket ban.
- Corrected classification
- Condemns. Schulman calls out feuding and dismisses the antitrust excuse. Sequencing a policy remedy before government involvement does not create an opposing assessment of the current failure.
We corrected the editorial classification of John Schulman's response about Labs’ failure to coordinate a concrete pacing proposal. Schulman calls out feuding and dismisses the antitrust excuse. Sequencing a policy remedy before government involvement does not create an opposing assessment of the current failure. This corrects our labeling, not a later change in the speaker's position.
This corrects an editorial classification.
Full primary public post and available surrounding thread. Speaker and position scope are explained in the linked event.
Classification correction
Correction to: Statement by Samuel Marks
- Previous classification
- Mixed or conditional. His personal-capacity statement supports the central criticism but presents continued safety work as a conditional rationale for remaining at Anthropic.
- Corrected classification
- Condemns. Marks endorses the concern, says robust alignment is unavailable and identifies himself with a slowdown demand. His safety job explains harm reduction; it is not endorsement of the current race.
We corrected the editorial classification of Samuel Marks's response about Unmanaged competitive development without robust alignment. Marks endorses the concern, says robust alignment is unavailable and identifies himself with a slowdown demand. His safety job explains harm reduction; it is not endorsement of the current race. This corrects our labeling, not a later change in the speaker's position.
This corrects an editorial classification.
Primary statement by an Anthropic safety researcher in his personal capacity. Marks supports Coxon’s central risk and coordination concerns while explaining why he continues safety work at Anthropic.
Share or embed this case
The embed keeps its attribution link to The Dispute Index. Credit the image to The Dispute Index and link to the case page.
Cite this record
- Publisher
- The Dispute Index
- Title
- Is Anthropic acting responsibly while advancing toward recursive self-improvement?
- First published
- Last updated
- Permalink
- https://disputeindex.com/cases/anthropic-responsibility-recursive-self-improvement-coxon
The Dispute Index. "Is Anthropic acting responsibly while advancing toward recursive self-improvement?". First published: 2026-09-10. Last updated: 2026-10-11. https://disputeindex.com/cases/anthropic-responsibility-recursive-self-improvement-coxon