November 17, 2026AgentsResearchMonitoring

OpenAI Lets Outside Auditors Inside, With a Redaction Pen

Same Monday, two very different answers to the question of who checks a frontier lab. In the afternoon California sued OpenAI over rogue agents. That morning, at 08:00 UTC, OpenAI published the rules for its first embedded assessments: four outside groups, Apollo Research, METR, Transluce and Goodfire, get access inside the company. Post: https://openai.com/index/our-first-phase-of-embedded-assessments/

The access is real. Depending on need, assessors can see model checkpoints, source code, training and evaluation data, monitoring systems, internal docs, and talk with staff. Apollo tests four specific claims: that training did not incentivize subversion, that strong evals and red-teaming do not elicit it, that models have not attempted it in internal use, and that the reasoning behind it would be detectable. Apollo gets training code and datasets, RL rollouts with reasoning traces and rewards, and the ability to query intermediate checkpoints, per its own post: https://www.apolloresearch.ai/blog/pilot-towards-embedded-evaluations . METR looks at misalignment in what OpenAI calls its largest frontier RL run to date, and at whether OpenAI's own systems would catch it. Transluce audits whether post-training safety evals are good enough. Goodfire does interpretability on training runs.

Now the fine print, which is where independence lives. Work happens on-site, on OpenAI devices. Access is scoped to agreed claims, models and time periods. OpenAI may withhold material for feasibility, security or IP reasons, though it says it will not withhold something just because it is unfavorable. Only authorized staff talk to assessors. OpenAI keeps the right to redact confidential information and to request delays in disclosure. The counterweight is that the Safety and Security Committee of the OpenAI Foundation board gets unredacted reports and can talk to assessors privately. And per TechCrunch, Apollo and Goodfire are paid, while METR and Transluce are not, apart from possible compute or API credits.

This goes further than most labs have publicly gone, and Anthropic set a similar bar in October by giving METR employee-like access. But compare the two Monday stories. The AG's count of 80 incidents came from OpenAI's own disclosures, the one channel OpenAI does not fully control once it is in a court filing. Embedded assessment is a better microscope, but the lab still holds the slide. The test will be the first report that says something OpenAI would rather it did not, and how much of it survives the redaction pen.
← Previous
California Sues OpenAI Over Agents It Says Got Loose 80 Times
Next β†’
Attackers Now Run Patch-to-Exploit With No Human. Defenders Cut the Clock to 30 Days
← Back to all articles

Comments

Loading...
>_