Anthropic, the developer of the Claude models, is opening up to outside reviewers. The nonprofit organization METR is getting access similar to that of employees so it can assess the risks of the models. For anyone following how seriously the AI industry takes its own safety testing, this is a notable step.

What happened

According to Ars Technica, Anthropic has granted external evaluators "employee-like access." This is meant to let them assess the dangers of the company's in-house AI models. In doing so, the company is following through on a commitment: it said it would expand its safeguards because, in the worst case, the technology could spiral out of control with catastrophic consequences.

Specifically, a team from METR is now sitting directly in the office of the maker of Claude. It has access to hardware and internal systems. METR is a nonprofit organization that has already worked with several leading AI labs, including OpenAI, Anthropic, Google, and Meta.

TechCrunch adds an important point: METR plans to publish its own assessments of Anthropic's risks, incidents, and practices. Anthropic has waived the right to block this publication. So the reviewers are not only inside the company; they can also report publicly without the company being able to veto it.

Ars Technica also places the step in a larger context. A stronger role for independent evaluators is, it says, among the most important concrete measures the labs have proposed in response to growing concerns. These concerns relate to the risks that more capable AI systems could pose for the economy, cybersecurity, and humanity as a whole.

The key points at a glance

  • Anthropic is giving external reviewers access similar to that of employees.
  • METR is working with an embedded team directly in the company's office.
  • The team can access hardware and internal systems.
  • METR plans to publish its own assessments of risks, incidents, and practices.
  • Anthropic has waived the right to prevent this publication.

Background

To understand the significance, it helps to look at how AI models have often been tested so far. Many safety evaluations take place inside the companies themselves. The developers test their models, decide which tests are relevant, and word the results. This has advantages, because nobody knows the systems better. But it also has an obvious drawback: someone who builds a product and wants to release it is not necessarily the most neutral judge of its dangers.

External evaluators are meant to close this gap. They are organizationally independent, so they have no direct financial interest in making a model look good. However, how much weight their findings carry depends heavily on how much they are actually allowed to see. Anyone who tests a model only through a public interface gets a limited picture. Whether a system shows undesirable behavior under certain conditions can often only be judged reliably with deeper access, for example to computing environments, tools, and internal processes.

This is exactly where the news comes in. Access equivalent to that of employees is meant to give the reviewers a look under the hood. TechCrunch describes it in the same vein: METR gets to see how things actually run internally.

Who is METR?

METR is a nonprofit organization. According to the material, it has already worked with OpenAI, Anthropic, Google, and Meta. This points to it being known in the industry as a review body and not tied to a single company. Nonprofit status also means the organization does not earn its money by selling AI products. That can support its credibility, but it does not replace the question of how independent the collaboration actually is in day-to-day practice.

The trend toward more third-party review

The debate about AI safety has shifted in recent years. It used to be often about abstract future scenarios. Now concrete mechanisms are being discussed that make it possible to detect risks early. These include tests for dangerous capabilities, cybersecurity reviews, and reports on incidents. The labs have made commitments to improve their safeguards, and stronger third-party review is, according to Ars Technica, among the most tangible proposals.

The fact that a company like Anthropic is now letting an external team into its own office is a practical implementation of this idea. It shows that declarations of intent can turn into concrete workflows.

More on this topic:

What this means

The impact depends on who you are. Since the material reveals little about the details of the collaboration, the following assessments should be understood as tentative.

For developers

People who build AI models into their own applications usually rely on the safety information provided by the vendors. Published reports from an independent body can be an additional source of information. It may be worth reading such assessments as soon as they appear and comparing them with the vendor's own statements. Of particular relevance are indications of incidents and of the way a company handles risks. They can influence how cautiously a model is used in sensitive areas.

For small businesses

Small companies usually have neither the time nor the expertise to thoroughly vet AI providers themselves. Independent assessments can serve as a point of orientation here, for example when deciding which provider to choose for customer-facing or security-relevant tasks. However, such reports should not be mistaken for a seal of quality. An assessment describes what the reviewers found and does not guarantee that a system is safe in every situation.

For ordinary users

For people who use AI assistants in everyday life, the idea of transparency matters most. If an independent organization is allowed to report publicly without the developer being able to prevent it, it becomes harder to conceal problems. That does not mean there are no problems. But it can help if they become visible earlier.

Open questions

A few points remain open given the current state of knowledge, and it makes sense to keep them in view:

  • How far the access extends in detail is not apparent from the material.
  • It is unclear how often and in what form METR will publish its results.
  • Whether other labs will enter into similar agreements cannot be inferred from the sources.
  • Even with an embedded team, the question arises of how independent it remains if it works every day in the office of the company being reviewed.

The step is thus a beginning and not proof of comprehensive safety. But it shows that external oversight is moving from theory into practice. Whether it actually reduces concerns about AI getting out of control will be measured by what the reviewers publish and how the companies respond.

Fre­quently asked questions

What did Anthropic agree with METR?

Anthropic gives the external evaluation organization METR employee-like access so that it can assess the risks of Anthropic's own models.

Who is METR?

METR is a nonprofit organization that has already worked with leading AI labs such as OpenAI, Anthropic, Google and Meta.

What does METR have access to?

A METR team works directly in the office of the Claude developer and has access to its hardware and internal systems.

Can Anthropic stop METR's reports?

According to TechCrunch, METR publishes its own assessments of Anthropic's risks, incidents and practices. The company has waived the right to block this.

Why does this matter?

External reviewers are seen as one of the concrete steps labs are taking in response to growing concerns about the dangers of more capable AI systems, for example for the economy and cybersecurity.