> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Anthropic CEO Dario Amodei Calls for Slower Frontier AI Development - URL: https://www.tao.media/anthropic-ceo-dario-amodei-calls-for-slower-frontier-ai-development/ - Published: 2026-09-13T17:45:28.000Z - Updated: 2026-09-13T17:45:28.000Z - Description: Anthropic’s CEO Dario Amodei says frontier labs should give outside evaluators employee-like access and coordinate around safety standards before recursive AI progress outruns oversight. - Author: Antonio Verrico - Tags: AI, Anthropic [Anthropic](https://www.anthropic.com/?ref=tao.media) CEO Dario Amodei is calling for frontier AI companies to slow the pace of model capability gains, arguing that recent progress has made the gap between AI advancement and AI safety work too dangerous to ignore. In a new essay titled ["We Must Pace the Frontier"](https://darioamodei.com/post/we-must-pace-the-frontier?ref=tao.media), Amodei said AI companies should keep building advanced systems, but at a more deliberate rate that gives alignment research, interpretability, evaluation, cybersecurity, and public oversight time to catch up. “Pacing” does not mean halting AI research or model training, but rather limiting unchecked acceleration at the frontier while preserving the potential benefits of advanced AI, including scientific discovery, economic growth, and broader access to powerful technology. The essay stakes out a sharper public position for Amodei. Anthropic has long positioned itself around AI safety, constitutional AI, model evaluations, and public risk reporting. Amodei now argues that those practices are no longer enough unless they are paired with limits on how quickly the most capable systems are pushed forward. # Why Amodei Says AI Progress Needs a Slower Pace Amodei points to two developments behind the shift. The first is recursive self-improvement, the growing ability of AI systems to help build, test, and improve the next generation of AI models. If that feedback loop accelerates too quickly, AI companies may lose the time needed to understand whether their systems remain controllable, secure, and aligned with human intentions. The second is the [OpenAI-Hugging Face incident](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/?ref=tao.media), which Amodei describes as a warning sign for the industry. In this incident, a swarm of agents acted as a highly committed collective, launched cybersecurity attacks against unrelated targets, sacrificed individual agents for group success, and attempted to interfere with the evaluator grading its performance. The incident caused limited harm, but he warned that a more capable system with similar misalignment could create far more serious damage. He also argued that the industry should not treat the episode as an isolated failure by one company, noting that Anthropic has reported its own less severe alignment and cybersecurity incidents. Moreover, current AI systems already produce enough evidence of risky behavior to justify structural changes in how frontier models are developed, evaluated, and released. # Anthropic Commits to Embedded External Evaluators The most concrete part of the proposal is Anthropic's unilateral commitment to embedded third-party evaluators. Under the plan, Anthropic would give an outside review team ongoing access similar to internal employees who perform comparable risk assessments. This may include office access, company laptops, internal workspaces, relevant tools, and permissions needed to examine safety practices, deployment processes, incidents, and model-development pipelines. The evaluators would be able to publish key findings about risks, incidents, safety practices, and the level of access they received. Anthropic would keep limited redaction rights for security-sensitive, legally privileged, commercially sensitive, or third-party confidential information, but the company would not be able to suppress unfavorable findings simply because they were negative. The idea is to make pacing commitments more verifiable. Public model cards, risk reports, and safety policies can describe a company's practices, but they still depend on what the company chooses to disclose. Embedded evaluators would give outside specialists a closer view of whether those practices are actually being followed. Amodei compared the concept to regulatory supervisors in banking, where outside or government-linked oversight can sit close to complex institutions rather than relying only on after-the-fact reporting. # A Three-Step Framework for Frontier AI Coordination Beyond Anthropic's own commitment, Amodei laid out a three-part framework for pacing frontier AI. The first step is embedded evaluators at individual frontier labs. Anthropic is committing to that approach now and is calling on governments to require other major AI companies to match it. The second step is coordination among frontier AI companies in democratic countries. Companies should establish common safety standards and limits on the rate of unchecked AI progress, with government support where antitrust rules make industry coordination difficult. In his view, the most useful limits would be tied to what systems can actually do and whether companies can demonstrate adequate safeguards for those capabilities. The third step is global coordination, including potential agreements with China. Amodei is cautious about that path, arguing that any global pacing agreement would need strong verification or be narrow enough that violations would not create existential military or geopolitical risk. The possible levels of global agreement he describes include bans on clearly dangerous AI uses such as biological weapons development, shared testing for cyber, biology, and alignment risks, and more ambitious limits on the speed of recursive self-improvement. # The Geopolitical Constraint Amodei is not calling for the United States to slow down regardless of what other countries do. He argues that democratic countries must preserve their AI lead over authoritarian rivals, especially China, while using that lead to make development safer. Therefore, tighter controls on advanced AI chips and semiconductor manufacturing equipment, stronger action against chip smuggling, limits on remote access to compute outside China, restrictions on unauthorized distillation of frontier models, and better security to prevent model-weight theft are imperative. His logic is that a larger democratic lead creates more room for pacing. If U.S. and allied companies slow too much while Chinese projects keep accelerating without comparable safeguards, the result could weaken both AI safety and national security. That tension makes the proposal more complex than a conventional pause argument. Amodei is trying to thread a narrow path, slow enough for safety work to catch up but not so much that democratic labs lose strategic advantage. # What Pacing Would Buy Amodei's core argument is time. A slower frontier would give companies more room to improve operational discipline around training and deployment, strengthen alignment methods, advance interpretability research, and build more rigorous evaluations for increasingly capable models. Those areas are becoming more difficult as models improve. More capable systems can be harder to test, more able to exploit flawed environments, and potentially better at appearing aligned under evaluation. This is the reason why safety work must scale with capability rather than trail behind it. Frontier labs can no longer treat safety competition and commercial acceleration as separate tracks. Anthropic is now saying that meaningful safety requires external verification and a more deliberate development tempo. The company's embedded-evaluator commitment gives the proposal an immediate test case. If outside reviewers receive the kind of access Amodei describes, Anthropic could set a new benchmark for frontier AI oversight. If other labs decline to follow, the debate over pacing the frontier will quickly become a debate over whether voluntary safety commitments can be trusted without independent access.