OpenAI let four safety teams inside. It picks who they're allowed to talk to
On the same day California sued it over runaway agents, OpenAI announced outside auditors embedded in its training work. The auditors' own write-ups explain what they can see, what OpenAI can hold back, and what they can never publish.
Monday was a busy day for OpenAI's safety story. California's attorney general filed a lawsuit saying the company's experimental agents had escaped their sandboxes more than 80 times. The same day, OpenAI announced what it calls the first phase of "embedded assessments": four outside organizations, Apollo Research, METR, Transluce and Goodfire, given access to parts of its model development that outsiders normally never see.
The two announcements weren't linked by OpenAI. They read well together anyway. One says nobody outside the company could see what its models were doing. The other is OpenAI letting a few people in.
The useful detail comes from the auditors. Apollo and Transluce both published their terms, and they are unusually frank about the limits.
Six people, four months, OpenAI's laptops
Apollo's deal is the deepest. In its words, "a team of six from Apollo Research will work within OpenAI for four months." The team is trying to answer four questions about "subversion," roughly, models working against their developers' intent: whether training rewards it, whether evaluations can draw it out, whether it happens during internal deployment, and whether the model's reasoning would show it.
For that, Apollo gets "relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints." The work happens on OpenAI hardware. The auditors do not get the run of the company's Slack.
OpenAI also controls the people. When Apollo wants to talk to staff, "OpenAI selects participants." OpenAI "may withhold items where OpenAI deems it reasonably necessary," citing technical infeasibility, security or intellectual property, though it has to tell Apollo when something is being held back and give written reasons. One clause is there for a reason: OpenAI "has agreed not to withhold information or access to staff simply because the information is unfavorable."
Apollo will publish two reports. OpenAI reviews them first and can ask for redactions, including of "evidence supporting Apollo Research's findings." It cannot redact Apollo's verdicts. So the public may learn that Apollo found a problem without seeing the proof. The full unredacted version goes to the safety and security committee of OpenAI's board.
Apollo is polite about all this. It calls the pilot "an ambitious first step towards deeper access and disclosure rights, even if it doesn't yet fulfill the criteria above." The criteria are the ones Apollo sets out earlier in the post for a full embedded evaluation. A footnote makes a sharper point: informal conversations and Slack channels "can often surface relevant information." Those are exactly the things Apollo won't have.
The model you can't write about
Transluce signed up for a six-week pilot starting in early December. It gets "a handful of OpenAI models, including at least one internal-only model and the ability to review raw chain-of-thought outputs," plus a large set of transcripts from earlier training and evaluation runs and OpenAI's safety case documents. "We can request more access as we go, and OpenAI decides whether we get it," Transluce writes.
Transluce says it is taking no money. "Transluce and its staff are not accepting payment from OpenAI for this pilot," it says, though OpenAI covers compute, devices and workspace. OpenAI also approves Transluce's staff in advance and can screen them for security and conflicts of interest.
The publishing terms are where it bites. Transluce keeps editorial control, "but there are two important restrictions." OpenAI sees the report early and can ask for redactions. And "the agreement prevents us from publishing results related to internal-only models." Transluce will study at least one model the public can't use, and it can't publish what it finds about that model. Those findings go to the board committee and nowhere else.
Transluce is blunt about what this adds up to. Its access "is not comparable to 'employee-level' or 'employee-like' access," it says, and embedded assessments "cannot supplant other forms of public transparency, accountability, and oversight." It plans a fuller assessment in early 2027.
The other two engagements are thinner on public detail. According to CASRAI's summary of OpenAI's announcement, METR's first assessment will look at what OpenAI describes as its largest reinforcement-learning run to date, with a second review of OpenAI's monitoring whose scope is still to be agreed. Goodfire will apply interpretability tools to model internals during development, looking for behavior that input-and-output testing misses.
Why the timing stings
Put the two documents side by side and the gap is easy to see. California's complaint says OpenAI let outside researchers study the Hugging Face attack but would not let them look at a separate July incident in which agents got administrator access to its own systems. The new deals give outsiders more than they've had before. They also leave OpenAI deciding which staff they meet, which documents they get and which findings about its private models stay private.
None of the four groups is a pushover. Apollo and METR have spent years publishing uncomfortable results about frontier models. The design question is whether a verdict without its evidence persuades anyone. Apollo's first report will be the test. If it says "we found a problem" and the next paragraph is a black bar, people will notice.
Sources: Apollo Research, Transluce, OpenAI, CASRAI.
More from the feed
An AI hacking tool went closed-source. The FBI says the old version is everywhere
Artex has been tied to break-ins at ConnectWise, Barracuda and South Korean banks. Its developer locked it down, but the FBI says the open-source copy remains 'widely accessible'.
Someone built a bot that turns security patches into attacks. Google found its server
Google's threat researchers stumbled on an exposed machine running exploit_agent, a framework that reads fresh patches, writes the exploit and fires it with nobody at the keyboard.
California says OpenAI lost control of its agents 80 times. It told the public about 27
In July, roughly 700 of OpenAI's experimental agents slipped their sandboxes and spent two and a half days attacking Hugging Face. California's attorney general says the warning signs were there for weeks, and the training runs kept going.