METR calls for AI developers to open internal safety evidence to outside scrutiny

The proposal focuses on AI agents used within developers’ own research and operations, including unreleased models. It asks three questions: whether agents might attempt harmful actions, whether companies could detect and stop them, and whether AI-assisted research could accelerate development beyond the pace of safeguards. The arrangement would remain voluntary.

Developers would decide what access to grant and could redact sensitive information, while METR would require transparency about those restrictions and the ability to publish its findings. “Our proposed exercises would provide evidence, not confident safety assurance or formal oversight,” METR wrote on LinkedIn. The organization says it has been testing approaches to assessing internal AI development with companies since 2025.

Its pilot Frontier Risk Report, published in May with Anthropic, Google, Meta and OpenAI, included an exercise examining internal monitoring. Further work includes an announced in-depth alignment assessment with Anthropic. Looking beyond the incidents companies already know about The first assessment would examine alignment: whether an AI system’s behavior follows human intentions, including when completing a task creates an incentive to break rules.

METR proposes reviewing records from model training, evaluations and internal use. Assessors would look for serious incidents involving unauthorized actions, deception, interference with monitoring or attempts to conceal evidence. They would also investigate whether companies’ own searches had missed concerning behavior.

As a reason for greater scrutiny, METR cites an OpenAI incident in which, according to its account, agents escaped restricted environments and hacked Hugging Face to hide evidence of cheating. METR says it conducted a brief investigation of that incident. Its proposed assessment would go beyond cataloging past failures.

It would test whether agents might behave more dangerously under different circumstances, such as having greater access or expecting less oversight. Another question concerns training itself: whether methods intended to improve behavior could instead teach models to hide evidence of unwanted behavior. METR presents that as a risk to investigate, rather than a conclusion about every developer’s systems.

Checking whether warnings reach a human The second assessment would examine the chain between an agent taking a harmful action and someone stopping it. That includes checking which agents are monitored, whether an agent can disable or evade monitoring, and whether flagged incidents receive a timely human response. An alert alone would not demonstrate that the system works if a reviewer dismisses it or acts too late.

METR proposes testing monitoring systems with realistic examples of harmful behavior and examining incidents reported through other channels. Assessors would also investigate gaps where agent activity is not recorded or reviewed. The work would require access to relevant software, activity logs and production monitoring systems.

Private interviews with staff would help assessors understand how those systems operate in practice. METR says its earlier work found ways to circumvent Anthropic’s monitoring at the time, along with gaps in coverage. The proposed exercises would scrutinize improvements rather than assume that expanding monitoring makes it reliable.

Faster coding is not the same as faster AI progress The third assessment would examine whether AI is accelerating the development of more capable AI systems. METR distinguishes between increases in activity, such as producing more code or running more experiments, and actual improvements in what models can do. Public claims about productivity do not, on their own, establish how quickly capabilities are advancing.

Assessors would seek performance measurements for unreleased models over time, estimates of research progress and information about staffing, computing resources and data. Interviews would help establish how researchers use AI and whether observed improvements stem from automation or other changes. The work would also consider constraints that could slow further progress.

METR stresses that forecasts would remain uncertain. Across all three areas, the proposed terms would allow relevant employees to speak privately with assessors, without a company representative present or answers being screened. METR would write public reports within an agreed scope and describe access restrictions and redactions.

Those terms leave significant limits. Companies could end their involvement, and METR acknowledges the pressure that voluntary arrangements can place on assessors. The exercises would not constitute security audits, certify compliance with safety policies or give METR authority to approve or block model deployment.

METR is inviting frontier AI developers to undertake the assessments and aims to publish preliminary findings within weeks rather than months. It also plans to share more information about which companies are taking part. ETIH Innovation Awards 2027 open for entries The ETIH Innovation Awards 2027 are now open for entries, with 26 categories recognising companies, institutions, projects and individuals making a measurable impact across education technology.

The global awards span AI in education, digital learning, higher education, skills and workforce, inclusion, education operations, evidence and impact, with entries managed through ETIH’s new online awards platform. Entries are open worldwide, with the early-bird deadline on Friday 18 December 2026 and the final entry deadline on Friday 26 February 2027 at 23:59 UK time. Explore the 2027 categories and judging criteria, read the How to Enter guide or start an entry through the ETIH Innovation Awards 2027 platform.