Artificial intelligence: Anthropic lets the first auditors into the company

As Anthropic reports in a blog post, the onboarding of six employees from the consulting firm METR, which specializes in AI models, has already begun. It is said that the auditors will not only be given Anthropic offices, but also company hardware and access to its internal networks. This means they have similar insights into current developments at Anthropic as the company's own employees.

Anthropic is drawing conclusions from a series of security incidents that have become known in recent months. At the end of July, the AI ​​pioneer had to admit that three other companies had been attacked while testing new models, which Anthropic employees only noticed during a subsequent check. These and other security lapses had fueled a global regulatory debate about the dangers of AI.

Dispute over transparency and trade secrets

The non-profit company METR investigated, among other things, the break-in of AI agents from Anthropic's competitor OpenAI at the AI platform Hugging Face. Its examiners are considered technically competent, but METR is accused of being very close to Anthropic.

As Anthropic announces, the company is also working with examiners from SecureBio and Faculty. There are also said to be discussions with other companies, including Hugging Face.

Anthropic boss Dario Amodei has long had plans for more control in AI development. The point of contention, however, was how transparent the auditors can be without disclosing business secrets. However, an insider told the Reuters news agency that the talks had recently progressed well and the company had become more open.

According to the regulation now presented, customer data, legally protected information and the model architecture are taboo for the auditors. Anthropic cannot veto METR's findings on risks and incidents, but can require narrowly defined redactions. METR may publicly challenge these redaction requests without disclosing the withheld information.

Anthropic outlines its three-tiered access system as follows: First-tier reviewers conduct narrowly scoped investigations and require permission to publish. In the middle level, the supervisors receive more information and can publish their findings independently, albeit with the redactions mentioned. The third and highest level, which is currently only intended for METR, offers employee-level access with similar publishing rights as the second level.