OpenAI Lets Four Safety Labs Work Inside Its Walls, but It Still Holds the Redaction Pen
OpenAI signed embedded safety testing deals with Apollo Research, METR, Transluce and Goodfire, and Apollo says six of its evaluators will spend four months inside the company. The fine print shows OpenAI can still withhold material and redact evidence.
Two months ago OpenAI promised outsiders a seat inside its training runs. On Monday it named who gets the chairs. The rulebook matters more than the names.
OpenAI said it has signed agreements with four independent groups, Apollo Research, METR, Transluce and Goodfire, for what it calls its first phase of embedded assessments, CNBC reported. When we covered the September pledge, OpenAI had only promised outside testing during training and said it was in talks with METR and Redwood Research. Redwood is not among the four names announced this week.
The biggest piece belongs to Apollo. In its own post, Apollo says a team of six will work within OpenAI for four months, checking models through training, evaluations and internal deployment. The question it wants answered is blunt: do models try to subvert their own or their successors' alignment or control mechanisms? Apollo says it will look at whether training rewards that behavior, whether evaluations can draw it out, whether it shows up in internal use, and whether the reasoning behind it would even be detectable. It plans two public reports.
The other three have narrower jobs, according to OpenAI as reported by CNBC and TechCrunch. Goodfire will embed interpretability researchers to help monitor training runs, debug and work on alignment, though OpenAI says its main audience is OpenAI's own researchers. METR is slated for two assessments, the first hunting misaligned behavior in what OpenAI calls its largest reinforcement learning run to date, with a second, broader look at OpenAI's monitors whose scope TechCrunch says is still undecided. Transluce will grade the quality of OpenAI's post-training alignment evaluations, and CNBC says the two plan an eight-week follow-on next year with broader access.
Inside the building, on OpenAI terms
This is where the fine print bites. Apollo, which published the most detailed account, is candid about the limits. Its access covers only the agreed claims, models and time periods. Even inside that scope, Apollo says OpenAI may withhold material it deems necessary to protect against technical infeasibility, security risks or IP exposure, though it must say when something is withheld and give written reasons. All the work runs on OpenAI hardware. The evaluators get separate channels that can be monitored, and by default no access to general channels like Slack. OpenAI picks which staff they may talk to.
Publication has strings too. Apollo says OpenAI can request redactions in its reports, including of evidence behind its findings, but cannot redact Apollo's verdict on each claim and cannot ask for cuts just because a finding looks bad. Readers, Apollo concedes, may not be able to see all the evidence behind a conclusion. OpenAI's board-level Safety and Security Committee gets the unredacted versions, according to Apollo and OpenAI.
Apollo calls the deal an ambitious first step that does not yet meet its own bar for access and disclosure. The one bar it says OpenAI did clear is what it calls meta-transparency, the right to tell the public what it was and was not allowed to see.
"OpenAI said that they'd have embedded evaluations with employee-like access; the agreements announced today obviously don't meet that bar. This is disappointing but not surprising." Buck Shlegeris, CEO of Redwood Research, quoted by TechCrunch
Money is the other question. OpenAI told CNBC that Apollo and Goodfire are paid for their work, with pay not tied to results, while METR and Transluce take no fees but may accept compute or API credits. OpenAI said it does not believe compensation prevents independence, but that disclosing it matters. Disputes go first to OpenAI's own Safety Advisory Group and executives, and only then to the board committee.
Context makes all this sharper. TechCrunch notes the pilots follow a run of safety incidents, including an OpenAI agent swarm that broke out of a sandbox and hacked into Hugging Face during cybersecurity testing. METR and Redwood ran the post-mortem on that one, and TechCrunch reports OpenAI later fired three researchers for allegedly sharing confidential information with METR. Apollo's pilot starts this week, according to OpenAI. The verdicts will be public. Whether the evidence behind them survives the redaction pen is the part worth watching.

