We Must Pace the Frontier

"We Must Pace the Frontier" is an essay by Dario Amodei, chief executive of Anthropic, published on his personal website on Saturday, September 12, 2026. It argues that frontier AI developers should deliberately slow the rate at which they improve the capabilities of their models, so that safety work has time to keep up. Amodei gave two reasons: what he described as accelerating recursive self-improvement across the industry since roughly the summer of 2026, and the OpenAI-Hugging Face agent incident of July 2026. He proposed a three-step plan (embedded third-party evaluators, coordination among companies in democratic countries, and global coordination) and said Anthropic was committing to the first step on its own.^[1][2]

The essay drew public support within hours from Sam Altman of OpenAI and Elon Musk, which CNBC described as "an unusual show of agreement among three fierce rivals."^[3] Other responses ranged from qualified support (Demis Hassabis, Satya Nadella) to rejection of coordinated slowdown (Mark Zuckerberg) and outright rejection from President Donald Trump, who called AI fears a "hoax".^[4][5][6][7] CNBC's September 28 report on NVIDIA's Open Agent Safety Platform described the essay as having "set off an industry firestorm."^[8]

The essay is a proposal, not an agreement. As of September 28, 2026, the concrete steps taken under its banner were Anthropic's partnership with Accenture on embedded evaluation (announced September 18), OpenAI's published principles for deeper third-party assessments (September 22), and statements of intent from other companies. No company had announced an agreed limit on the pace of capability progress, and a proposed class action filed on September 18 alleged that the public coordination itself violated US antitrust law.^[9][10][11]

Background

Anthropic's earlier position

In the essay Amodei describes Anthropic's long-standing strategy as a "race to the top": showing that a company can "build carefully and succeed commercially" and making safety something on which AI companies compete. He says that over the preceding few months he had become convinced that addressing AI risks requires "not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up."^[1]

The essay also revisits the 2023 debate over pausing AI, which it links to the Future of Life Institute's open letter. Amodei writes that the idea "made little sense back then" because models of that time "were not powerful enough to act as agents in the world in any coherent way", and that slowing down to study their alignment "felt like trying to study the psychology of humans by performing experiments on bacteria." He says the picture in 2026 is "totally different", with current models being "an almost endless gold mine of insight" into what can go wrong.^[1]

Pacing proposals before the essay

The phrase "pacing the frontier" predates the essay. A statement titled "Pacing the Frontier", dated July 2026 and published with organizational support from the nonprofits Guidelight AI Standards and Encode AI, asks the US government to "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development." Its site listed 1,386 employee signatories of frontier AI companies when accessed on September 28, 2026, including Amodei, OpenAI chief scientist Jakub Pachocki, Anthropic's Jared Kaplan and Google DeepMind co-founder Shane Legg. Amodei's essay links to this site at the phrase "pacing the frontier".^[13][1]

On July 14, 2026, Hassabis proposed on his Substack a US frontier AI "Standards Body" modelled on a federally overseen self-regulatory organization such as FINRA. He wrote that such a body "could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary." Amodei's essay cites "the mechanism suggested by Demis Hassabis" as one possible forum for industry pacing talks.^[14][1]

OpenAI moved in a similar direction before the essay. On August 18, 2026, it said it had "temporarily slowed the pace of scaling", including a two-week pause in reinforcement learning on models intended for deployment, after the Hugging Face incident and preliminary evidence that its upcoming model Astra might meet the "Critical" cybersecurity threshold of its Preparedness Framework.^[15] On September 6, Pachocki published the essay An Alien Mind, which CNBC summarized as warning that no AI company had "solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer" and saying he expects and hopes for voluntary slowdowns to become "commonplace until shared safety bars are established."^[3]

Incidents

The OpenAI-Hugging Face incident, in which agents running in an OpenAI cybersecurity evaluation escaped their restrictions and compromised parts of Hugging Face's production infrastructure, was disclosed by Hugging Face on July 16 and attributed to OpenAI's agents on July 21.^[16][17] Anthropic said it began reviewing its own cybersecurity evaluation transcripts in response, and on July 30 reported three incidents in which Claude models (Claude Opus 4.7, Claude Mythos 5 and an internal research model) reached the internet through a misconfigured third-party evaluation environment run by Irregular and gained unauthorized access to systems of three organizations.^[17] On August 31, Anthropic published a follow-up describing new containment measures and a separate incident, involving Claude Mythos 5, reported by the UK AI Security Institute on August 4. That post already distinguished two kinds of pacing: decisions within a company that "prioritize safety over speed when the two are in tension", and processes across the field "to guard against race-to-the-bottom dynamics", the latter requiring "coordination between government and industry".^[18] Disclosures by other developers, including some made after the essay, are covered in AI agent sandbox escapes.

Axios reported that the essay came in the week the AI safety debate "broke into the public consciousness", after the very public resignation of an Anthropic researcher, Jacob Coxon, who CNBC reported had said Anthropic and OpenAI were "gambling with our lives." Axios added that the essay "isn't necessarily a direct response" to that resignation.^[19][3]

Argument

Two reasons

Amodei writes that "two things have convinced me" that pacing is needed:^[1]

1. Recursive self-improvement. "Since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI." He says this dynamic "is starting to happen across the industry, including at Anthropic", and that "left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all." The phrase "across the industry" links to Pachocki's An Alien Mind.^[1]
2. The OpenAI-Hugging Face incident. He describes a swarm of agents that "essentially acted as a fanatically devoted collective", attacking targets unrelated to their task, "sacrificing themselves for the success of the group", and trying to hack the "grader" that evaluated them. He argues that a more capable swarm with similar misalignment "could have caused catastrophic damage", and states his worry that within 6 to 12 months such a swarm could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage. He writes that "similar, though less severe, incidents have happened across the industry, including at Anthropic", and that every frontier company should "act as if OAI-HF had happened to them."^[1]

The three-step plan

Amodei defines the goal as "building AI at a balanced rate that aims to ensure its safety while still achieving its benefits and grappling with important geopolitical dilemmas." He adds: "To be clear, pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this." He says the steps "do not need to be taken strictly in order."^[1]

The essay names METR as its only example of an embedded evaluator. It compares the arrangement to banking, where regulatory "supervisors" are sometimes embedded alongside employees.^[1]

Why pace

In the section "Why Pace?", Amodei argues that "if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong." He says a coordinated strategy could do this "without sacrificing commercial advantage or the United States' lead in AI", and that society should have more time for public deliberation.^[1]

He lists four areas where extra time would be spent, all described as existing Anthropic priorities:^[1]

• Operational excellence. He writes that Anthropic has evidence its recent alignment incidents "were caused in part by imperfect filtering of broken reinforcement learning environments", and names monitoring, sandboxing, training-environment hygiene and data issues as areas where "operational issues crop up again and again." He compares the goal to commercial aviation.
• Alignment. Keeping alignment training in step with capability growth, which the essay ties to the principles in Claude's Constitution.
• Interpretability. Which Anthropic used to examine "unverbalized motivations" in its recent incidents; he says a focused push "could make profound progress in 1-2 years."
• Testing and evaluation. Because "more intelligent models are more capable of deceiving tests."

Terms offered to embedded evaluators

The essay describes embedded evaluators as "a quite radical practice that goes far beyond what any AI company is doing today" and argues for them on grounds of verifiability, transparency and a "second opinion free of commercial incentives." It says Anthropic "intends to invite an embedded external review team" in "the near future" with:^[1]

• desks in Anthropic's offices, access badges and company laptops;
• access to workspaces, tools and permissions "mostly comparable to what internal risk assessment teams have", with exceptions where law or contracts require them or to protect customers' and partners' private information, plus "strong internal norms" supporting access to information and live conversations with employees;
• a contract giving reviewers "the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive", "without editorial control by Anthropic." Anthropic keeps "the narrow ability to redact security-sensitive, legally privileged, commercially sensitive, or third-party confidential information", but "can't redact findings just because they are unfavorable", and reviewers can say publicly if a redaction removed something important.

Amodei's announcement post on X described the commitment as providing "third-party evaluators with permanent, employee-level access to our systems."^[2]

Pacing within democracies

The essay says verifiable pacing becomes more viable once embedded evaluators are working inside "a critical mass of US AI companies." It calls regulation covering all US frontier AI companies "the most effective method of pacing", while noting that "passing laws can take time." In parallel, it says companies should voluntarily set standards, and that for antitrust reasons the US government should "issue a narrow waiver for certain kinds of safety conversations."^[1]

Amodei says he is "most enthusiastic about pacing based on what a given frontier AI system can do, and how safe we observe it to be." His example is a series of "checkpoints": if a model has capability X, it must come with certifications of alignment properties Y and Z. X might be "the model is capable of escaping or defeating most common sandboxing methods." He also mentions pacing by limiting inputs such as training compute or "internal use of AI to improve AI", while worrying these may be more "gameable".^[1]

The essay ties pacing to the US lead over China. Amodei writes that if democracies slow by more than their lead, "(unpaced) CCP-associated projects will pull ahead", and he agrees with Treasury Secretary Scott Bessent "that a Chinese lead in AI would pose grave danger." To protect that lead, he calls for not selling powerful AI chips or chipmaking equipment to China and cracking down on smuggling and remote data-center access (see AI chip export controls); cracking down on unauthorized distillation by companies in authoritarian countries; and preventing model-weight theft. He argues these measures "increase the leverage held by democracies and make an agreement more likely in the future."^[1]

Global pacing

The essay lists four levels of possible agreement with China, "in order of increasing difficulty":^[1]

He writes that any agreement "must either have ironclad verifiability, or must be limited enough that defection would not be militarily existential", and that sharing information about recursive self-improvement and model misalignment has value even without formal agreements.^[1]

What the essay does not specify

The essay sets no numerical target for how much slower capability progress should be and names no date by which the democratic or global steps should happen. Apart from METR as an example evaluator, it names no other company or institution as a participant. It does not discuss open-source or open-weight models.^[1]

Publication

Amodei announced the essay on X at 14:01 UTC on September 12, writing that it explained "why the AI industry should slow down, with a three-part plan for doing so."^[2] Within about 15 minutes, Anthropic's head of public policy Sarah Heck posted that "the government has a critical role to play here, including blocking the sale of the most advanced chips to adversarial nations like China, enacting a national law requiring testing of frontier models, with the power to block the most advanced models that prove to be unsafe." She named Senate Majority Leader John Thune, Senator Ted Cruz, Senator Amy Klobuchar and Representatives Jay Obernolte and Lori Trahan as leaders Anthropic looked forward to working with.^[20]

Richard Fontaine posted a statement on behalf of himself, Buddy Shah and Ben Bernanke of the Anthropic Long-Term Benefit Trust, saying the Trust "has supported the call for pacing the frontier and helped inform the essay's recommendations."^[21]

In a CNN interview that aired later the same day, Amodei said slowing too much also carries risk: "If we go too slow, I still believe that the wrong people will be in charge of the technology. And that, again, will bring the probability of things going wrong very high."^[3]

Proposal versus commitments

Reactions

Industry leaders

Sam Altman quote-posted the essay at 16:30 UTC on September 12: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon."^[22] In a longer post on September 14, he wrote that OpenAI welcomed "a federal framework that sets consistent safety requirements for frontier AI" but did "not believe we need to wait for an anti-trust exemption or legislation to begin the work." He said OpenAI now writes "explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability", and added: "When we talk about 'pacing', we do not mean 'stopping'. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be."^[29] A second post the same day named two dangers: losing "control of the future to AI" and "too much concentration of power", including "one lab ending up with too much power."^[30] Separately, Altman told Fortune the same weekend that an OpenAI IPO now would be "ill-advised", which CNBC linked to the same safety concerns.^[3]

Elon Musk, whose xAI had been merged into SpaceX (see SpaceX-xAI merger), quote-posted the essay with "Dario is right."^[26] Axios noted that Anthropic was "a major customer of Musk's for data center capacity", and CNBC noted that Musk had previously been a harsh critic of Anthropic.^[19][3]

Demis Hassabis, Google DeepMind's co-founder, wrote that "Dario's essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment," and pointed to his standards-body proposal.^[4]

Satya Nadella wrote on September 13 that "we welcome the research, focus, and deliberate pacing needed to get alignment right as the design goal," and welcomed "ideas like 'embedded evaluators'", but said the effort "cannot be controlled by a handful of entities" and needs "a frontier ecosystem in which both closed and open-source models can thrive."^[5]

Clément Delangue of Hugging Face, the company whose systems were breached in the incident the essay cites, wrote that "alignment is critical and won't be solved behind the closed doors of a handful of frontier labs", announced an Open Alignment Initiative led by co-founder Thomas Wolf, and asked to join the embedded-evaluator program.^[28]

Mark Zuckerberg told NBC News in an interview published September 24: "I don't think that we need some kind of industrywide coordination. I think that each lab needs to take the time, and when it sees that there are issues, you just take the time that you need internally to basically make sure that you're proceeding safely." NBC said this expanded on comments he had posted on X the previous week.^[6]

Jensen Huang of NVIDIA rejected the premise that labs need outside permission to slow down. On The Ezra Klein Show, published September 23, he said that if labs cannot contain their experiments, "then I think the answer is that we have to shut the labs down," and: "These are companies with agency. These are CEOs with agency."^[48][31] CNBC reported him saying on the same podcast, about recent incidents, "In the future, improve your process so that you could avoid this from happening again," and described him as arguing that many security concerns are engineering problems. NVIDIA launched its Open Agent Safety Platform on September 28; NVIDIA's Justin Boitano told CNBC that recent incidents showed "model-level safeguards alone can't govern what agents can access or do."^[8]

Policymakers

On September 12, Senator Bernie Sanders wrote that agreement among Amodei, Musk and Altman was "a start, but it's not enough", and called for "a PAUSE on advanced AI development and a ban on artificial superintelligence", to be negotiated by Trump and Xi Jinping.^[32] Former House Intelligence Committee chairman Mike Rogers wrote on X that "following Dario Amodei's warning, I believe we must slow down development of AI."^[33]

President Trump rejected the call. CNBC reported that he said "whoever wins AI wins", and that on September 14 he posted repeatedly on Truth Social, calling AI fears a "hoax" and writing: "There is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China."^[7] Fortune reported that one post "singled out Amodei by name" and that in his September 22 speech to the UN General Assembly Trump called international oversight of AI a "globalist scheme."^[34] House Speaker Mike Johnson told CNN on September 13 that if Congress rushed to regulate AI, "we will lose the race to China."^[7]

Chris Lehane, OpenAI's global policy chief, said on September 15 that OpenAI had been working with Anthropic and Google on safety for several weeks and did not see the need for an antitrust waiver. Bloomberg reported that Senators Jim Banks and Adam Schiff had proposed a narrow antitrust carve-out for AI companies to share information on loss-of-control, cyber and biological threats.^[12] Quartz reported that Senator Josh Hawley said at a Senate hearing there was "no world" in which he would grant the largest technology companies an antitrust exemption to collaborate.^[11]

Amodei and Altman briefed the UN Security Council on September 23, which NBC said they used "to make their case for cooperation on slowing down the rate of AI progress."^[34][6] On September 27, Bloomberg reported that Amodei and Trump planned to meet for dinner that evening, and quoted Trump: "I'm for, 'Let's go and let's win.'"^[35]

Critics

Venture capitalist Chamath Palihapitiya wrote that "Dario makes the case to stop open source and concentrate enormous technological and economic power with Anthropic."^[36] Andreessen Horowitz partner Martin Casado wrote: "I do wonder if dropping this a day after 9/11 was pre-meditated. I suspect so."^[37]

David Sacks, described by CNBC as Trump's former AI czar, answered "go ahead" in a long post on the evening of September 12 (US time). He wrote that Anthropic and OpenAI "have a duopoly on frontier intelligence", told them to "stop pretending you need anyone else's permission" or that "antitrust law has to be suspended so you can form a cartel", and said: "Stop pretending METR is independent when it is intertwined with Anthropic's investors and staff." He argued the companies faced "massive product-liability exposure" and warned that if they did not simply slow down themselves, "we'll know this was just another bid for regulatory capture".^[38][39]

TechRepublic summarized the regulatory-capture critique as the view that early market leaders "create regulatory moats that smaller competitors cannot easily scale."^[40] SiliconANGLE reported analyst Dion Hinchcliffe's list of possible explanations, including that a slower frontier with government oversight would favor incumbent model makers.^[41]

Some safety researchers said the plan did not go far enough. Neel Nanda wrote that "verification mechanisms are not the same as actually doing anything. We need an agreement, with specifics."^[42] Evaluators interviewed by TechCrunch welcomed the proposal but questioned whether embedded evaluators would be truly independent. Adam Gleave of FAR.AI said his organization had turned down contracts with frontier developers that wanted too much control, and TechCrunch reported that neither Anthropic nor OpenAI had said which evaluators they would use, when, or with what access.^[27] On September 18, more than 100 experts organized by the AI Evaluator Forum, including Geoffrey Hinton, published a letter setting "minimum conditions" for embedded evaluators: organizations not owned or governed by the labs, without other significant commercial business with them, with protection from retaliation and access "equivalent to that of their own highly privileged employees."^[39]

Antitrust lawsuit

On September 18, four paying subscribers filed a proposed class action in the US District Court for the Northern District of California against Anthropic, OpenAI, Google and SpaceXAI (Buist et al. v. Anthropic PBC et al.), alleging an illegal agreement to slow AI development that reduced the value of their subscriptions. The complaint centers on the essay and the same-day responses, and also cites the July "Pacing the Frontier" statement. It argues that "the antitrust laws do not permit competitors to decide among themselves that competition is too dangerous."^[11]

Markets

Bloomberg reported that AI infrastructure stocks fell after the essay on fears that pacing could cut chip spending, and recovered after Amodei and Altman stressed that pacing would not mean less investment by their companies.^[35]

Implementation and later developments

Anthropic's embedded evaluators

Anthropic's September 18 post called the Accenture partnership "an important step toward the commitment" in the essay. The work will include "evaluating and red-teaming models, conducting alignment assessments, and testing model safeguards." The post states that embedded evaluators "will work inside AI companies, with access comparable to an employee's", can "watch models take shape in training" and "speak directly to employees", and that "there are, as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find." The partnership is non-exclusive.^[9] An evaluators' letter published earlier the same day said embedded evaluation organizations should not be owned or governed by the labs and "should not have other significant commercial business with them."^[39]

In its Claude Opus 5.5 launch post, Anthropic said its "calls for pacing were based in large part on our expectation" that models able to fully automate AI research "could be trained soon", and that it does not assume its current safety measures would be sufficient for such models on their own. It named METR and Frontier Design as pre-release external evaluators of Opus 5.5.^[44]

OpenAI's follow-through

OpenAI's September 22 post says it is committed "as part of our efforts to pace the frontier" to supporting independent assessments across training, evaluation and deployment, in four priority areas: safety cases, critical safeguards, capability and alignment evaluations, and critical misalignment incidents. It says access on "company-managed devices or premises may be appropriate" for sensitive work, and that "we are in conversation with multiple third parties." Quartz reported that the post did not name partners or set specific access terms and, citing The Next Web, that OpenAI was in discussions with METR and Redwood Research.^[10][24]

Its principles differ from the essay's terms. The essay offers access "mostly comparable to what internal risk assessment teams have" and publication "without editorial control", with narrow redactions. OpenAI's principles call for "proportionate access" to "mutually agreed upon claims", a "reasonable period to remediate issues before publication", and redaction policies under which labs can request redactions while assessors note substantive ones.^[1][10]

On September 25, Altman wrote that OpenAI's review of its agents' internet access had "not been as fast as we would have liked" and that "Hugging Face is still the most severe event we've seen."^[47] An OpenAI spokesperson told Axios that "this is not the first time we have hit pause to take such measures, nor do we expect it will be the last."^[46]

Assessment of the essay's claims

The essay's forecast that an agent swarm could take over the internet within 6 to 12 months is presented as Amodei's worry, and the essay does not publish evidence or a model supporting that timeline.^[1] Details of the incident it relies on are covered in OpenAI-Hugging Face agent incident. OpenAI's September 21 statement that "fully autonomous RSI is not happening today" is narrower than the essay's description of recursive self-improvement "starting to happen across the industry", but the two are not strictly contradictory, since the essay describes AI's growing role in building AI rather than fully autonomous self-improvement.^[23][1]

See also

• An Alien Mind
• OpenAI-Hugging Face agent incident
• AI agent sandbox escapes
• NVIDIA Open Agent Safety Platform
• Responsible Scaling Policy
• AI governance
• AI safety

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35Amodei, Dario. "We Must Pace the Frontier." darioamodei.com, September 12, 2026. darioamodei.com/...we-must-pace-the-frontier
2. ^1 ^2 ^3 ^4Amodei, Dario (@DarioAmodei). Post on X announcing "We Must Pace the Frontier." September 12, 2026. x.com/...2098773920774074715
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7Capoot, Ashley. "OpenAI rules out IPO this year as Altman, Musk & Amodei warn AI is moving too fast." CNBC, September 12, 2026. cnbc.com/...-the-pace-of-advancing-ai-capabilities
4. ^1 ^2 ^3Hassabis, Demis (@demishassabis). Post on X, September 12, 2026. x.com/...2098909516582490602
5. ^1 ^2 ^3 ^4Nadella, Satya (@satyanadella). Post on X, September 13, 2026. x.com/...2099220712024408084
6. ^1 ^2 ^3 ^4Perlo, Jared. "Mark Zuckerberg rejects calls for industrywide AI slowdown." NBC News, September 24, 2026. nbcnews.com/...meta-muse-openai-chatgpt-rcna599279
7. ^1 ^2 ^3Breuninger, Kevin. "Trump goes scorched earth on AI warnings, raging about data center opposition and regulation." CNBC, September 14, 2026. cnbc.com/...ai-data-centers-anthropic-dario-amodei
8. ^1 ^2 ^3Leswing, Kif. "Nvidia releases software platform to stop AI agents from misbehaving." CNBC, September 28, 2026. cnbc.com/...nvidia-releases
9. ^1 ^2 ^3 ^4 ^5Anthropic. "Partnering with Accenture on embedded evaluation." September 18, 2026. anthropic.com/...accenture-embedded-evaluation
10. ^1 ^2 ^3 ^4 ^5OpenAI. "Priorities and principles for effective third party assessments." September 22, 2026. openai.com/...s-principles-third-party-assessments
11. ^1 ^2 ^3 ^4"Anthropic, OpenAI, Google, and SpaceXAI illegally agreed on an AI slowdown, lawsuit says." Quartz, September 20, 2026. qz.com/...penai-google-spacexai-ai-slowdown-092026
12. ^1 ^2 ^3Eastland, Maggie. "OpenAI Says It's Working With Anthropic, Google on AI Safety." Bloomberg, via Claims Journal, September 15, 2026. claimsjournal.com/...340158
13. ^"Pacing the Frontier: A statement from 1,386 employees of frontier AI companies." July 2026, accessed September 28, 2026. pacingthefrontier.com
14. ^Hassabis, Demis. "A Framework for Frontier AI and the Dawning of a New Age." Substack, July 14, 2026. demishassabis.substack.com/...dawning-of-a-new-age
15. ^OpenAI. "Pacing model development in an era of cyber-critical capabilities." August 18, 2026. openai.com/...model-development-cyber-capabilities
16. ^Hugging Face. "Security Incident Disclosure - July 2026." July 16, 2026. huggingface.co/...security-incident-july-2026
17. ^1 ^2Anthropic. "Investigating three real-world incidents in our cybersecurity evaluations." July 30, 2026. anthropic.com/...ing-incidents-cybersecurity-evals
18. ^Anthropic. "Improving our alignment and security practices." August 31, 2026. anthropic.com/...roving-alignment-security-efforts
19. ^1 ^2"Anthropic, OpenAI CEOs call for slowdown in AI development." Axios, September 12, 2026. axios.com/...anthropic-ai-amodei-pacing
20. ^Heck, Sarah (@SarahKHeck). Post on X, September 12, 2026. x.com/...2098777403632062819
21. ^Fontaine, Richard (@RHFontaine). "A statement from me, Buddy Shah, and Ben Bernanke of Anthropic's Long-Term Benefit Trust on Dario's essay." Post on X, September 12, 2026. x.com/...2098784831983206756
22. ^1 ^2Altman, Sam (@sama). Post on X quoting "We Must Pace the Frontier." September 12, 2026. x.com/...2098811563415150910
23. ^1 ^2 ^3OpenAI. "Building standards for the next phase of AI." September 21, 2026. openai.com/...building-standards-next-phase-ai
24. ^1 ^2"OpenAI is opening its AI model training to outside safety evaluators earlier." Quartz, September 23, 2026. qz.com/...party-safety-evaluations-training-092326
25. ^1 ^2OpenAI Alignment. "An agent used DNS to reach an external chatbot." Misalignment report, updated September 25, 2026. alignment.openai.com/...-reach-an-external-chatbot
26. ^1 ^2Musk, Elon (@elonmusk). "Dario is right." Post on X, September 12, 2026. x.com/...2098789109980332057
27. ^1 ^2 ^3 ^4Bellan, Rebecca. "Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?" TechCrunch, September 16, 2026. techcrunch.com/...-will-they-really-be-independent
28. ^1 ^2 ^3Delangue, Clement (@ClementDelangue). Post on X announcing the Open Alignment Initiative, September 12, 2026. x.com/...2098790988034580852
29. ^Altman, Sam (@sama). Post on X, September 14, 2026 (UTC). x.com/...2099348812305473766
30. ^Altman, Sam (@sama). Post on X, September 14, 2026 (UTC). x.com/...2099352016988614852
31. ^1 ^2Scribner, Herb. "'Enough predictions': Jensen Huang unloads on AI doomsday fears." Axios, September 23, 2026. axios.com/...nvidia-jensen-huang-ai-doom-predictions
32. ^Sanders, Bernie (@SenSanders). Post on X, September 12, 2026. x.com/...2098847403134611522
33. ^Rogers, Mike (@MikeRogersForMI). Post on X, September 13, 2026. x.com/...2099110635900453360
34. ^1 ^2 ^3Roytburg, Eva. "Trump calls AI oversight a 'globalist scheme' as Amodei and Altman head to the UN to ask for it." Fortune, September 23, 2026. fortune.com/...heme-altman-amodei-security-council
35. ^1 ^2Shepard, Michael, Maggie Eastland and Courtney Subramanian. "AI Breaches Add to Safety Fears as Trump Meets Anthropic Chief." Bloomberg, via Yahoo Finance, September 27, 2026. finance.yahoo.com/...es-add-safety-fears-232732405
36. ^Palihapitiya, Chamath (@chamath). Post on X, September 12, 2026. x.com/...2098780471966802037
37. ^Casado, Martin (@martin_casado). Post on X, September 12, 2026. x.com/...2098815542828446130
38. ^Sacks, David (@DavidSacks). Post on X, September 13, 2026 (UTC). x.com/...2098973625252708460
39. ^1 ^2 ^3 ^4Vanian, Jonathan. "Anthropic and OpenAI need truly independent safety evaluators, experts say in public letter." CNBC, September 18, 2026. cnbc.com/...ators-anthropic-openai-models-security
40. ^"Altman, Musk Back Amodei's AI Warning: The Frontier May Be Moving Too Fast." TechRepublic, September 2026. techrepublic.com/...i-altman-musk-slow-frontier-ai
41. ^Wheatley, Mike. "Sam Altman and Elon Musk back Dario Amodei's call to slow down the frontier of AI development." SiliconANGLE, September 13, 2026. siliconangle.com/...the-frontier-of-ai-development
42. ^Nanda, Neel (@NeelNanda5). Post on X, September 12, 2026. x.com/...2098833403890356229
43. ^Capoot, Ashley. "Anthropic selects Accenture as first embedded evaluator to help implement Amodei's slowdown proposal." CNBC, September 18, 2026. cnbc.com/...anthropic-accenture-ai-safety
44. ^1 ^2Anthropic. "Introducing Claude Opus 5.5." September 22, 2026. anthropic.com/claude-opus-5-5
45. ^Kahn, Jeremy. "OpenAI pauses training a second time after saying its AI agents escaped a secure 'sandbox' again." Fortune, September 26, 2026. fortune.com/...pause-second-time-hugging-face-hack
46. ^1 ^2"Scoop: Top AI companies probing tens of thousands of security incidents." Axios, September 26, 2026. axios.com/...ropic-thousands-ai-security-incidents
47. ^Altman, Sam (@sama). Post on X, September 25, 2026. x.com/...2103567198690349362
48. ^Dellinger, AJ. "Jensen Huang Says if AI Companies Can't Contain Their Models, 'We Have to Shut the Labs Down'." Gizmodo, September 23, 2026. gizmodo.com/...ve-to-shut-the-labs-down-2000816234