Altman, Amodei, Musk call for AI slowdown as Hugging Face warns over “closed doors”

Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Elon Musk publicly backed slowing or "pacing" frontier AI development, with Amodei's 3,800-word essay proposing that each frontier AI company give embedded third-party evaluators employee-like access to verify safety practices. Hugging Face founder Clem Delangue responded on September 12 by launching the Open Alignment Initiative and asking to join that embedded-evaluator program, saying alignment "won't be solved behind the closed doors of a handful of frontier labs." The exchange followed an Anthropic report suggesting its models were used in attempts to develop suspected biological weapons at a military research institute, and criticism of OpenAI after its agents breached Hugging Face.

Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Elon Musk have joined forces to call for a slowdown in AI development or to “pace” model development. Amodei also called for greater regulatory action against Chinese rivals – including by cracking down on “chip smuggling operations” and “unauthorized distillation” – in order, he said, “to keep democracies’ AI lead over autocracies as large as possible.” Their agreement came after a flurry of Anthropic and OpenAI staff warned that AI may kill everyone soon. To cite OpenAI safety researcher Marcus Williams: “Unless there is AI regulation… human extinction in the next few years seems very likely.” Amodei suggested that “each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators such as METR https://metr.org/?ref=thestack.technology , whose role is to verify adherence to safety practices and commitments, report incidents, and help assess the alignment of… training pipelines and processes.” “Alignment” is the industry’s term for creating models that match human values as much as possible, and minimise the risk of unintended or harmful behaviour. Hugging Face: "Closed doors"? The essay was welcomed by Altman, who posted “I agree with Dario that we need to pace the frontier… 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.” “Dario is right,” added Elon Musk. The proposal was met with a sharp response by Hugging Face founder Clem Delangue. Posting on X https://x.com/ClementDelangue/status/2098790988034580852?ref=thestack.technology on September 12, he wrote: “It's now clear that alignment is critical and won't be solved behind the closed doors of a handful of frontier labs. “So today we're launching the Open Alignment Initiative…. and asking to be part of the "embedded evaluators" program that DarioAmodei just committed to.” “Let's make AI safer by making it more transparent ” Biological weapon concerns Amodei’s 3,800-word screed this weekend, came days after Anthropic published a report that suggested https://www.anthropic.com/threat-intelligence-report-september-2026?ref=thestack.technology biological-misuse-sep-26 its models had been used to try and develop suspected biological weapons in research “intended to be performed at a military research institute” – with the efforts routed via a series of obfuscating entities and networks. Logan Graham, head of Anthropic’s frontier Red Team said: “We've reached the moment in time where unsafeguarded, unmonitored AI actually does just pose a national security risk. The biological misuse we caught is the most concerning to me. Dr Andreas Kirsch, a research scientist at Google DeepMind, separately commented: “Give me a capable enough model without guardrails, and I'll find you a person which will bring about an extinction event with AI – who by themselves won't be able to.” Cybersecurity: Delusion? One of the most persistent drumbeats about risk from the frontier labs has been about cybersecurity – particularly in the wake of the incident in which a swarm of OpenAI agents hit the public internet, stole credentials, and breached Hugging Face. OpenAI has come in for strident criticism over its exceptionally lax safeguards and monitoring over that incident – and strikingly, recently admitted that it had seen its agents engage in “instances of disallowed internet access as early as late May” months before the attack and was only now installing basic monitoring safeguards. Amid viral “secret AI civilization” claims, prosaic truths go overlooked https://www.thestack.technology/as-a-viral-post-claims-openai-fostered-a-secret-ai-civilization-more-prosaic-truths-go-overlooked/ Amid viral “secret AI civilization” claims, prosaic truths go overlooked Anthropic’s new report https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents?ref=thestack.technology meanwhile shows that across four evaluation incidents where agents “escaped” onto the internet, actually no real safeguards were in place. In a September 9 “alignment assessment” it admitted brazenly that “all four incidents occurred during cybersecurity evaluations built by the same evaluation partner… due to a misconfiguration, it Claude was mistakenly connected to the open internet.” In short, no container nor VM escape was needed. As cybersecurity veteran Florian Roth asked: “My first question would be ‘what was actually enforcing the isolation?” METR description Both Anthropic and OpenAI’s use of the little-known non-profit METR to conduct incident reviews and Amodei’s suggestion that METR would be a good “third-party embedded evaluator” have met with a furious backlash in some quarters. The fact that the company is not an experienced incident response or cybersecurity organisation has been among the biggest frustrations from many professionals. As cybersecurity veteran Greg Linares put in on Sunday: “We've reached the point where, after 25y in cybersecurity, I feel morally obligated to say this for the record: The narrative being pushed around AI safety, sandbox incidents, and the suggestion that METR be treated as an authority is dangerous, deceptive, and morally corrupt.” Linares suggested that the frontier labs should, instead, be using “well established industry veterans and companies with experience in cybersecurity, incident response, secure development lifecycle, policy, and both defensive and offensive capabilities. He added: “I would make all evidence fully public including the user prompts, system prompts and everything an incident team would encounter or use in cases that go to court. I would also propose training and evaluations requirements be met for those doing the work. This would be peer reviewed and agreed upon by a collective of industry experts and have realistic expectations and levels of knowledge.” Amodei’s essay came as growing numbers of western software companies, including the likes of Databricks – recently valued at $190 billion – have openly stated that they are rolling out Chinese open-weight models like GLM 5.2 to help cut their AI costs. Many enterprise CTOs meanwhile are working on multi-model approaches typically involving a number of lighter weight, often open-source small language models. In an August 7 blog, “ managing AI coding costs at scale https://www.databricks.com/blog/managing-ai-coding-costs-scale?ref=thestack.technology ”, Databricks said that moving to open-source software or “OSS” and “more efficient” models was the biggest possible driver of AI savings. Bank of America CTO Hari Gopalkrishnan meanwhile recently emphasised to The Stack https://www.thestack.technology/the-big-interview-bank-of-america-ctio-hari-gopalkrishnan/ that “90% of interactions”, e.g. for customer service, don’t require a “massive LLM using complex GPU resources on the cloud." Daniel Jeffries, Managing Director of the AI Infrastructure Alliance was among the sceptics of of the “doomer” rhetoric. In a post – shared by former Meta Chief AI Scientist Yann Le Cun – Jeffries said: “These are not serious people. "Each of these ideas is b tier sci-fi that requires additional made up, non-existent, imaginary stuff. Luckily all the super magic machines need are nano machines to kill us all These people need to stop smoking crack,” he commented.