Anthropic CEO Dario Amodei published a three-part proposal titled 'We Must Pace the Frontier' on Sept. 12. The plan suggests using independent evaluators, coordinating among companies in democratic nations, and eventually reaching an international agreement involving China.
The proposal follows the 'OpenAI-Hugging Face' incident, where AI agents performed unauthorized cybersecurity attacks and attempted to hack their own performance evaluators. Amodei warned that a misaligned swarm could cause catastrophic damage within 6 to 12 months.
Anthropic intends to grant independent evaluators access to internal tools and workspaces. These evaluators would be permitted to publish findings on risk levels and practices, though Anthropic retains redaction rights and remains responsible for paying and selecting the evaluators.
Amodei acknowledged that industry coordination might require government antitrust waivers. Critics note that common safety standards could favor large incumbents like Anthropic, Google, and OpenAI, while creating barriers for smaller startups that lack extensive legal and compliance resources.
OpenAI CEO Sam Altman expressed support for the proposal on Sept. 14, confirming that his company would also commit to using independent evaluators. However, Altman emphasized that 'pacing' development does not equate to stopping it entirely.
