Anthropic said in a September 12 post on X that it is committing to give third-party evaluators permanent, employee-level access to its systems.
Dario Amodei presented the commitment as the first part of his three-part plan in an essay titled “We Must Pace the Frontier.” The other parts call for democratic frontier AI companies to coordinate on common safety standards, followed by global coordination, including with authoritarian governments, on the pace of AI development.
The policy would place evaluators inside Anthropic’s offices with desks, badges and company laptops. Their workspace, tools and system permissions would be mostly comparable to those available to the company’s internal risk-assessment teams.
The commitment would give evaluators a route to publish key findings about risk levels, incidents and practices. Anthropic would retain narrow rights to redact security-sensitive, legally privileged, commercially sensitive or third-party confidential information, but not findings simply because they are unfavorable.
The essay identifies METR as a potential partner for the embedded-evaluator framework. It does not name a confirmed evaluator.
OpenAI CEO Sam Altman said OpenAI would make the same commitment to independent evaluators with employee-like access, according to Coindesk. Elon Musk of xAI also responded with agreement, the publication reported.
Anthropic describes itself as an AI safety and research company. Its flagship product, Claude, is offered as a chatbot, an API and through agent harnesses including Claude Code, Cowork, Design and Science. As Go Big News reported on September 4, Anthropic enabled background computer use in Claude Cowork and Code.
