
OpenAI lets four outside labs inside to test for scheming, with limits
By Yash Malviya·
Apollo Research, Goodfire, METR and Transluce will probe OpenAI's models for deception on site. Their own posts show what OpenAI can redact and what it cannot.
On November 16, 2026, OpenAI announced a program to let outside researchers work inside the company to test whether its models deceive, evade monitoring or pursue goals they were not given. Four groups are involved: Apollo Research, METR, Goodfire and Transluce. The most useful reading is not OpenAI's announcement, which we could not open, but the terms the evaluators themselves published. Those terms are specific, and they set limits.
What Apollo will do
Apollo Research's post, "A Pilot Towards Embedded Evaluations," says six of its staff will work inside OpenAI for four months. They will assess model alignment across training, evaluations and internal deployment, and investigate whether models try to subvert their own or their successors' alignment or control mechanisms. The four questions are whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether the related reasoning is detectable.
Access is partly settled. For training incentives, Apollo gets relevant training code and datasets, reinforcement-learning rollouts with reasoning traces and rewards, and the ability to query intermediate checkpoints. Access for the other claims is still to be agreed. Work happens on OpenAI hardware, on monitored channels, without default access to general channels such as Slack.
What Transluce will do
Transluce, a nonprofit lab, will review OpenAI's internal post-training alignment evaluations to see how well they detect deception, manipulation and reward hacking. Access includes a handful of models, one of them internal-only, raw chain-of-thought outputs, select non-public evaluations and training transcripts. Transluce says the access is "not comparable to 'employee-level' or 'employee-like' access." The pilot lasts six weeks from early December, with a broader assessment planned for early 2027. According to The Neuron's account, Goodfire will examine what happens inside models during training, and METR will look at OpenAI's largest reinforcement-learning run to date.

The independence terms
The rules on who controls what are the heart of this. Both posts describe the same architecture:
“OpenAI has agreed not to withhold information or access to staff simply because the information is unfavorable.”
- OpenAI may withhold in-scope items for technical infeasibility, security risk or IP protection, but must tell Apollo and give written reasons.
- OpenAI may request redactions of confidential information. It may not redact Apollo's verdicts, or redact merely because a finding is unfavorable.
- Unredacted reports go to the Board's Safety and Security Committee.
- Apollo can describe its scope, what it requested and received, what was withheld and any pressure it believed shaped its work, without OpenAI's prior review. It calls these "meta-transparency" rights a "minimum standard for evaluation engagements."
- Transluce has editorial control, must submit its report for IP redaction, and can note when restrictions materially affected its conclusions.
- OpenAI approves the embedded staff and screens for security and conflicts.
Apollo says plainly that the pilot "does not yet meet" its full access and disclosure criteria, and calls it a first step. That candor is the most credible sentence in the whole package.
The gap the program is trying to close
The need is visible in earlier work. In a post published October 22, 2026, METR reported on a three-week pilot, July 21 to August 11, in which a METR and a Redwood Research employee stress-tested OpenAI's internal misalignment monitoring. They found, among other things, a gap that left roughly 10 percent of eligible internal Codex traffic unmonitored from May to July, an exploit that could keep hundreds of recent actions unmonitored, and two client-controllable headers that could disable monitoring for new agent instances. OpenAI reported that most issues were patched. METR did not verify that. Our earlier piece on the monitoring report and the changed patch wording tracks that.
That exercise was brief and redacted. The new program is longer and gives the evaluators more rights to describe what they could not see. It is an improvement in kind, though still short of the independence a regulator would ask for.
Questions the first reports should answer
The posts make the evaluation criteria easy to write down in advance. When Apollo's two reports and Transluce's report appear, a reader can check them against the published terms:
- Did OpenAI withhold items, and did it give written reasons by category and approximate volume, as Apollo's terms require?
- Did OpenAI decline any attestation of claims drawn from employee conversations, and what did that do to the conclusions?
- Did redactions affect conclusions, a statement both Apollo and Transluce reserve the right to make?
- For Transluce, does the final report document operating conditions under the AEF-1 standard from the AI Evaluator Forum, as it says it will?
- What happened when an access request was escalated through OpenAI's point of contact, executives and the Safety and Security Committee?
A report that answers all five is worth far more than one that reaches a clean verdict without them. The terms were designed so that silence on any of them would itself be informative.
One more point on scope. The posts describe pilots, not permanent arrangements. Apollo calls its engagement a first step, Transluce plans a fuller assessment only in early 2027, and METR's further work is described by The Neuron as under consideration. Nothing yet commits OpenAI to any particular level of outside access beyond these pilots.
Our take
This is a real, modest step. Four months inside a lab beats a few days of API access, and the right to disclose that something was withheld matters more than the right to see everything. But the evaluators are partly funded, in some cases, or supported with compute by the company they assess, and results on internal-only models stay private. The reports will tell us more than the announcement does. Read them for one line above all: what OpenAI withheld, and why.
Frequently asked questions
What are OpenAI embedded assessments?
A program announced November 16, 2026 in which outside groups, including Apollo Research and Transluce, work inside OpenAI to test models for deception, evasion of safety systems and unassigned goals.
Which organizations are involved in OpenAI's embedded assessments?
Apollo Research, METR, Goodfire and Transluce. Apollo and Transluce published their own terms on November 16; the roles of METR and Goodfire come from The Neuron's account.
Can OpenAI hide unfavorable findings?
Per Apollo, OpenAI agreed not to withhold information or staff access simply because it is unfavorable, and may not redact Apollo's verdicts. It can still withhold items for security or IP reasons, with written reasons.
How long do the OpenAI assessments last?
Apollo's pilot is four months. Transluce's pilot is six weeks starting in early December, with a broader assessment planned for early 2027, details not finalized.
Will Transluce publish results on OpenAI's internal models?
No. Transluce says results on internal-only models cannot be published, so public findings will cover only released models. Full unredacted findings go to the board's Safety and Security Committee.
Sources
What each one is, and whose it is.
- 1
A Pilot Towards Embedded Evaluations, Apollo Research (November 16, 2026)
OtherIndependent of the vendor - 2
Embedded assessment pilot with OpenAI, Transluce (November 16, 2026)
OtherIndependent of the vendor - 3
Red-teaming OpenAI's agent monitoring, METR (October 22, 2026)
OtherIndependent of the vendor - 4
OpenAI outsiders catch AI deception, The Neuron (November 17, 2026)
Press reportIndependent of the vendor