Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

AI news & research

News ·

OpenAI lets four outside groups work inside its labs to audit safety

OpenAI has signed agreements with Apollo Research, METR, Transluce and Goodfire for embedded assessments with access to checkpoints, training data, monitoring systems and staff. METR will look for misaligned behaviour in OpenAI's largest reinforcement learning run to date, Apollo at whether models covertly work against their developers, and Transluce at alignment tests. OpenAI can redact confidential details and ask for publication delays.

Outside safety testers usually see an AI model near the end, through an API, under time pressure. OpenAI is now letting four groups work inside the company, with access close to what its own researchers have, during training and internal use. The first phase covers Apollo Research, METR, Transluce and Goodfire.

Each has a different brief. METR will look for misaligned behaviour in OpenAI's largest reinforcement learning run to date, including failures that existing tests and monitors may have missed, and will set its own priorities within the agreed scope. Apollo Research will test whether models could covertly work against their developers' intentions and whether OpenAI's safeguards would catch it, assessing four claims in two stages. Transluce will review the quality and coverage of OpenAI's post-training alignment evaluations, with a broader follow-on of about eight weeks planned for early 2027. Goodfire will embed interpretability researchers to examine model internals, mainly for OpenAI's own researchers, while publishing what it learns.

Access can include unreleased models and checkpoints, training data, monitoring records and staff. Assessors handling sensitive material work on site on OpenAI devices. Where weights or architecture are too sensitive, OpenAI may pair assessors with its own researchers instead. The Safety and Security Committee of the OpenAI Foundation board gets unredacted reports, can speak to assessors privately, and settles disagreements that the company's Safety Advisory Group and executives cannot.

The context is the White House Accord on Super Intelligence, under which labs agreed to independent assessment of their safety controls. Anthropic gave METR similar employee-like access in October, and METR's 11-week review there has already been published. OpenAI's version is broader in the number of groups involved, and it builds on the monitoring stack it described last month.

The weak points are in the fine print. OpenAI keeps the right to redact confidential information and to ask for disclosure delays. Conversations with staff are off the record; assessors can only publish statements OpenAI agrees to attest, or note that it declined. Conflict-of-interest and cool-off rules are still being written, and a broader METR monitor review is pending agreement on scope. Independence will be judged by what the first reports actually say, and how much of them is blacked out.

Our take: This is deeper access than most labs have offered, and it follows Anthropic's similar arrangement with METR. The weak points are OpenAI's right to redact and staff interviews held off the record, so the published reports will show how independent the process really is.

Source: OpenAI

More stories