Anthropic allows METR researchers to conduct on-site AI safety assessments

Article

Anthropic lets 6 outside researchers conduct on-site AI safety reviews: reports require no approval, but the company retains the right to redact or delete content

Beating AI News reported that Anthropic announced it had allowed independent AI evaluation organization METR’s 6 researchers to enter the company, providing office space, work computers, and system access broadly comparable to that of the internal safety assessment team. They will no longer merely test Claude externally, but will also be able to access internal tools and investigate dangerous model behavior, safety incidents, and safeguards. Anthropic compared this arrangement to banking regulators assigning resident examiners to banks.METR can publish risk assessments and investigation findings independently, and Anthropic has no right to prevent publication of the reports. However, when customer data, legally privileged materials, or model architecture are involved, the company can still require the relevant content to be deleted. If METR considers the redactions or deletions unreasonable, it may publicly object, but it cannot disclose the deleted original text.

Anthropic divides external evaluations into 3 tiers, and METR has entered the highest-permission Tier. The nonprofit SecureBio is reviewing its biochemical weapons risks, while Accenture’s Faculty is also advancing model red-teaming and safety assessments.METR explicitly stated that it does not accept compensation or donations from frontier AI companies, while Anthropic had previously announced direct funding for Accenture’s evaluation work.METR acknowledged in a statement published on October 14 that the current cooperation still depends on AI companies voluntarily providing information; evaluators may not obtain key evidence, and vendors can influence what the public sees by restricting access and redacting or deleting report content.

Original link https://m.theblockbeats.info/flash/372203

Key points

Anthropic announced that it would allow independent AI evaluation organization METR’s 6 researchers to enter the company and provide office space, work computers, and system access broadly comparable to that of the internal safety assessment team.
The article states that METR researchers can access internal tools and investigate dangerous model behavior, safety incidents, and safeguards.
METR can independently publish risk assessments and investigation findings, and Anthropic has no right to prevent publication of the reports.
When customer data, legally privileged materials, or model architecture are involved, Anthropic can still require the relevant content to be deleted.
If METR considers the redactions or deletions unreasonable, it may publicly object, but it cannot disclose the deleted original text.
Anthropic divides external evaluations into 3 tiers, and METR has entered the highest-permission Tier.
SecureBio is reviewing Anthropic’s biochemical weapons risks, while Accenture’s Faculty is advancing model red-teaming and safety assessments.
METR stated that it does not accept compensation or donations from frontier AI companies; the article states that Anthropic had previously announced direct funding for Accenture’s evaluation work.
METR In October 14 a statement, it said that the current cooperation still depends on AI companies voluntarily providing information; evaluators may be unable to obtain key evidence, and vendors may influence the public content by restricting access and redacting or deleting reports.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is equivalent to allowing external safety inspectors to enter AI the company’s internal operations. Compared with testing Claude only from outside the company, on-site researchers can access more tools and information and can publish investigation findings independently. However, this does not mean they can see all materials: the company can still restrict access and require the deletion of content involving customer data, confidential materials, or model architecture.

Why it matters to readers

On-site assessments may bring external organizations closer to model safety issues and incident evidence and improve the independence of evaluation findings. However, the company retains some control over access and redactions or deletions, so the reports ultimately seen by the public may remain incomplete.

Ordinary users can use this to assess AI whether the company’s security assessments are conducted solely internally, and how much independent investigative and public-disclosure latitude external evaluation bodies actually have.

Risks and unknowns

The provided materials do not include an Anthropic announcement or METR a complete explanation, so the specific scope of permissions cannot be independently verified.
It remains unclear which systems, incident records, and key evidence researchers can actually access.
The specific procedures, decision criteria, and actual use of Anthropic’s requests to delete or revise reports are unknown.

Anthropic allows METR researchers to conduct on-site AI safety assessments

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center