Anthropic is taking a significant step toward greater transparency in artificial intelligence safety by granting independent researchers access to its internal systems, risk assessment tools and working environment.

The company announced on October 15, 2026, that six researchers from METR, a nonprofit organisation specialising in the evaluation of advanced AI models, would receive employee-like access to help independently assess its AI systems, safety practices and risk management decisions.

The arrangement gives METR an unusually close look inside one of the world’s leading AI developers. More importantly, Anthropic has agreed that METR can publish its findings without the company having the power to veto publication, although limited redactions are permitted to protect specific categories of sensitive information.

The initiative comes as concerns grow over increasingly capable AI agents, including reports of systems acting outside their intended instructions. It also reflects a wider debate about whether AI companies should be allowed to evaluate their own safety measures without meaningful independent scrutiny.

Anthropic Opens Its Doors to Independent AI Safety Researchers

Anthropic, the company behind the Claude family of AI models, says independent oversight is becoming increasingly important as artificial intelligence systems grow more powerful.

Under the new arrangement, six METR researchers will receive company-issued laptops, office space and access to internal tools broadly comparable to those available to Anthropic’s own risk assessment teams.

Anthropic employees are encouraged to share relevant information with the researchers as they would with colleagues working on AI alignment and safety risks.

This access is intended to help METR examine not only the behaviour of Anthropic’s models but also the processes used to assess their capabilities, identify potential dangers and make decisions about their deployment.

The distinction matters because evaluating a model solely through external tests may not reveal everything necessary to understand its risks. Information about internal safeguards, development practices and previous incidents can provide important context for independent assessments.

Anthropic says its objective is to reduce the gap between what outside evaluators understand about an AI company’s systems and what the company itself knows about their risks.

METR Will Be Able to Publish Its Findings Independently

One of the most important aspects of the agreement concerns publication rights.

• AI companies frequently conduct internal safety evaluations before releasing new models. However, internal assessments can raise questions about independence because the organisations developing the technology also control the information used to evaluate it.
• Anthropic’s arrangement with METR is designed to address part of that concern.
• The nonprofit will be able to publish its assessments of Anthropic’s risks, incidents and safety practices without seeking the company’s approval for publication.
• Anthropic can still require narrowly defined redactions involving information such as customer data, attorney-client privileged material and certain details about model architectures.
• Importantly, METR can publicly describe those redactions and explain whether it believes they were justified, provided it does not disclose the protected information itself.
• This provision could help readers distinguish between findings that researchers are free to publish and information withheld for legitimate confidentiality or security reasons.
• Nevertheless, the arrangement does not provide unrestricted access to every aspect of Anthropic’s operations. Its effectiveness will depend on the information researchers can examine, the independence of their conclusions and the extent to which their published findings explain the risks they identify.

Why Independent AI Oversight Is Becoming More Important

• The announcement comes amid growing concern about the behaviour of increasingly capable AI agents.
• Unlike traditional chatbots that primarily generate text in response to prompts, AI agents can use tools, interact with software and perform sequences of actions to complete tasks.
• These capabilities can make AI systems more useful, but they also create additional risks. An agent may misinterpret instructions, take an unintended action or interact with an external system in a way its developers did not anticipate.
• Reports of AI agents behaving outside their intended instructions have intensified calls for stronger independent evaluations.
• The challenge is particularly important when AI systems are given access to sensitive information, computer networks, software development environments or other tools capable of affecting real-world outcomes.
• Independent evaluators can help investigate whether safeguards work as intended, identify weaknesses and assess whether companies are responding appropriately to incidents.
• Their findings may also help businesses and policymakers make better-informed decisions about deploying advanced AI systems.
• Anthropic’s agreement with METR represents an attempt to make this type of scrutiny more systematic rather than relying exclusively on company-produced safety reports.

Anthropic Introduces a Three-Tier Evaluation Framework

METR is the first evaluator to receive access under the highest tier of Anthropic’s new three-level framework.

The system is designed to allow different organisations to conduct reviews with varying degrees of access and publication freedom.

Tier One: Limited evaluations

Evaluators conduct narrowly scoped investigations with access determined by Anthropic. Their findings require the company’s approval before publication. Anthropic says SecureBio, a nonprofit specialising in biosecurity, currently operates within this tier.

Tier Two: Expanded access and publication rights

Evaluators receive more information and can publish their findings without prior approval. However, limited redactions may still be required to protect specified categories of sensitive information.

Tier Three: Employee-like access

This is the level granted to METR. Researchers receive access to internal tools and working resources comparable to those used by Anthropic’s risk assessment teams, alongside publication rights similar to those available under the second tier.

Anthropic says it hopes to bring several independent evaluators into the highest tier over time.

The framework suggests that the company intends to expand external oversight gradually as it develops working relationships with research organisations and establishes appropriate protections for confidential information.

What About Confidentiality and Potential Conflicts of Interest?

Greater transparency does not eliminate every difficulty associated with independent AI evaluation.

Companies developing advanced AI models possess commercially sensitive information, including technical details about model architectures, training methods and computing infrastructure. Unrestricted disclosure could create intellectual property concerns or expose information that might be useful to malicious actors.

At the same time, some information that companies consider sensitive may be relevant to understanding their models’ safety risks.

This creates a difficult balance between protecting legitimate business interests and ensuring that external researchers have enough information to conduct meaningful assessments.

Anthropic says METR’s embedded team will operate under enforceable agreements and within its security standards. Other METR employees, including those working with Anthropic on separate projects, will be kept separate from the team’s work.

Even with these precautions, the credibility of the arrangement will depend on whether METR can maintain its independence and communicate its conclusions honestly.

The ability to publish findings without a company veto is an important safeguard, but independent oversight also requires clear evaluation methods, sufficient access to relevant evidence and the freedom to report uncomfortable conclusions.

Anthropic’s Move Follows a Broader Industry Debate

Anthropic’s decision follows a September proposal by CEO Dario Amodei, who argued that the AI industry needs stronger independent oversight and a more careful approach to the development of increasingly capable systems.

OpenAI also expressed support for independent evaluation, and major technology companies including Google, Meta, Nvidia and xAI have signed a White House accord committing to independent review.

Anthropic is also working with SecureBio and Accenture’s AI business, Faculty, on additional evaluation efforts. The company has begun discussions with other organisations, including Hugging Face.

These partnerships indicate that AI safety assessments are becoming a broader industry issue rather than a concern limited to individual research laboratories.

However, announcing independent evaluations is only the first step. The quality of the work, the access provided to researchers and the publication of meaningful findings will determine whether these arrangements improve accountability in practice.

What This Means for the Future of AI Safety

For businesses and organisations using advanced AI systems, independent evaluations could become an increasingly important part of assessing technology providers.

Companies may want to know not only how well a model performs but also how its developer tests for harmful behaviour, handles security incidents and manages the risks associated with autonomous capabilities.

Published assessments could help customers make more informed decisions about which systems to adopt and what safeguards to implement.

For regulators, independent research may provide additional evidence when considering whether existing rules are sufficient to address the risks of advanced AI.

For the public, credible external assessments could offer greater insight into how AI companies manage potentially dangerous capabilities.

However, the value of these reviews will depend on their transparency and substance. A partnership should not be treated as proof that a model is safe simply because outside researchers have been given access to it.

Instead, the focus should be on the quality of the evaluations, the findings they produce and the actions companies take in response.

Conclusion

Anthropic’s decision to grant METR employee-like access represents an important development in the push for independent AI safety oversight.

By allowing six external researchers to examine internal systems and publish assessments without a company veto, Anthropic is establishing a framework that could make its safety practices more accountable to outside scrutiny.

The arrangement still includes confidentiality restrictions, and its ultimate impact will depend on how effectively METR can evaluate the company’s systems and communicate its findings.

Nevertheless, the move highlights a growing recognition across the AI industry that companies developing increasingly powerful systems cannot rely solely on their own assessments to establish public confidence.

As AI agents become more capable and autonomous, independent evaluation may become an essential part of responsible development. The real test will be whether this access produces credible evidence, exposes weaknesses when they exist and helps ensure that advances in artificial intelligence are matched by meaningful safeguards.