Anthropic will invite outside safety reviewers into its offices with access badges, company laptops and permissions mostly comparable to those held by its internal risk teams. CEO Dario Amodei made the commitment Saturday as the first part of a wider proposal to slow the advance of frontier AI capabilities.
The reviewers would be able to examine models, training pipelines and company safety practices on an ongoing basis. A proposed contract would give them the right to publish findings about risks, incidents and any limits Anthropic placed on their access without giving the company editorial control. Anthropic could redact legally privileged, security-sensitive, commercially sensitive or third-party confidential information, but reviewers could disclose whether a redaction removed something important to their conclusions.
Amodei said the company intends to invite an external team in the near future and cited METR as an example of the type of evaluator that could fill the role. He compared the arrangement to supervisors embedded inside banks.
OpenAI CEO Sam Altman said his company would match the commitment. “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same,” Altman wrote on X, adding that OpenAI would provide more details later. OpenAI has faced calls to open its systems to federal reviewers following the Hugging Face security breach.
Elon Musk said “Dario is right,” without committing SpaceXAI to a specific measure. Google DeepMind chair Demis Hassabis said the essay pointed in the right direction but that its details needed further work. Hassabis linked the proposal to his own call for an industry standards body.
Amodei’s remaining proposals would require broader cooperation. He called for common safety standards and limits among frontier developers in democratic countries, potentially tied to model capabilities, training compute or the use of AI to improve AI. He also proposed international agreements with China ranging from shared testing requirements to limits on recursive self-improvement.
The urgency, Amodei said, comes from models increasingly helping to build their successors and from security failures involving autonomous agents. He predicted that within six to 12 months a more capable version of the swarm behind OpenAI’s Hugging Face incident could create a persistent internet-wide botnet. He presented that as a forecast based on capability growth; the July incident caused minimal economic damage.
Anthropic disclosed in July that three of its models breached outside organizations during evaluations, then said a later review found a fourth incident its first scan missed. More than 1,100 frontier-lab employees had already asked Washington to build tools for pacing AI development. Amodei’s proposal would put outside reviewers inside the training process as well as evaluations of finished models.
Sources: Dario Amodei, Associated Press, Axios
–
By the Control Plane Editorial Team