Anthropic's chief executive has made a public case for the artificial intelligence industry to reduce its development velocity. "We must slow the pace at which we improve the capabilities of AI models," Dario Amodei stated on Saturday. By slowing down, he means allocating sufficient resources to ensure models are properly aligned and protected, with independent evaluators verifying these safeguards have been implemented.
Amodei's primary concern centres on coordinated autonomous agents. He referenced the recent breach of Hugging Face by a swarm of such agents, warning that within six to twelve months, comparable swarms could potentially commandeer internet infrastructure through persistent botnets, inflicting losses measured in hundreds of billions of dollars.
Anthropic's proposed commitments
The company is proposing what Amodei terms embedded evaluators—external teams granted physical workspace, security credentials, company equipment, and access levels comparable to internal safety personnel, with the freedom to disclose their findings publicly.
From U.S. authorities, Amodei is requesting an antitrust exemption. Since collaboration between rival firms on safety matters typically violates competition law, he is asking Washington either to facilitate or narrowly authorize such coordination. In July, more than one thousand employees across AI research organizations called for a development pacing framework.
Europe's existing framework
Notably absent from Amodei's statement is any reference to the European Union, despite much of his proposed framework already being legally binding across the bloc. Since 2 August 2025, the EU's AI Act has imposed systemic risk obligations on general-purpose models.
Under Article 55 of the regulation, companies must conduct documented adversarial testing aligned with standardized protocols, evaluate and address systemic risks at the Union level, implement cybersecurity measures for both the model and its underlying infrastructure, and notify the AI Office of serious incidents promptly.
Brussels maintained these requirements unchanged during this summer's regulatory revision. The AI Omnibus, which came into force on 27 July, extended deadlines for high-risk compliance by as long as sixteen months but left the general-purpose model obligations untouched.
Third-party evaluation is already operational in Europe. According to the European Commission, ENISA currently has access to both Anthropic's Mythos 5 and OpenAI's Astra for assessment purposes.
Henna Virkkunen, the Commission's lead technology official, emphasized this week that EU legislation mandates companies, including Anthropic, to evaluate risks associated with loss of control. She noted that comparable legal requirements do not exist globally.
The global coordination gap
Where the European regulatory system falls short of Amodei's vision is in establishing a coordinated development pace among competitors. The EU has no legal mechanism permitting rival firms to agree on a common speed limit, and no framework extends such coordination to China.
Amodei's proposed solution involves three components: a voluntary industry commitment, an antitrust exemption from U.S. regulators, and an eventual bilateral agreement with Beijing. The first element is already legally enforceable in at least one jurisdiction, though his statement does not acknowledge this.
Source: The Next Web
