OpenAI gives four outside audit organizations on-site access to its internal AI systems for safety reviews
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
OpenAI is giving Apollo Research, METR, Transluce and Goodfire direct on-site access at its offices to unreleased models, source code, training data and monitoring systems, plus confidential conversations with developers. Apollo examines covert misalignment against developers' intentions, METR reviews misbehavior in the largest RL training run, Goodfire analyzes internal model processes, and Transluce evaluates post-training safety checks. An independent safety board receives unredacted reports and mediates disputes. Auditors publish findings independently, but OpenAI can redact confidential material and delay reports to fix risks. An eight-week OpenAI-Transluce follow-up is planned for early 2027.
Entities: OpenAI, Apollo Research, METR, Transluce, Goodfire, White House Accord on Super Intelligence
0 primary
What happened
OpenAI says it is giving four outside groups (Apollo Research, METR, Transluce and Goodfire) on-site access at its offices to unreleased models, source code, training data and monitoring systems, plus confidential talks with developers. Apollo will look for covert misalignment, METR will review misbehaviour in the largest reinforcement learning training run, Goodfire will analyse internal model processes, and Transluce will evaluate post-training safety checks. An independent safety board gets unredacted reports and mediates disputes. Auditors publish on their own, but OpenAI can redact confidential material and delay reports to fix risks. An eight-week OpenAI-Transluce follow-up is planned for early 2027.
Why it matters
If it works as described, outside researchers could check claims about training and monitoring that were previously taken on trust, which matters to regulators, enterprise buyers and rival labs setting their own standards. It also gives OpenAI a way to show compliance with its stated commitments under the "White House Accord on Super Intelligence". The real value depends on the redaction and delay powers: OpenAI controls what the public sees and when, so independence is partial. The announcement comes from OpenAI itself, with no published contract terms, so how much access is real is not yet verifiable.
What is noise
"First time" independent on-site access is OpenAI's framing, and it is not independently confirmed. The announcement gives no scope, duration, staffing, or funding details, and it does not say who pays the auditors, which bears on independence. OpenAI's own expectation of "early friction" and talk of reviewing company culture are intentions, not commitments.
Watch next
- 01Whether Apollo, METR, Transluce and Goodfire publish their own findings, how long after review, and how much is redacted or delayed by OpenAI.
- 02The early 2027 eight-week OpenAI-Transluce follow-up: does it happen on schedule and produce a public report?
- 03Statements from the auditors or the safety board confirming the access terms, who funds them, and any disputes the board has to mediate; also whether rival labs or regulators adopt similar arrangements.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More regulation signals
Full feed →- NYT-led publishers file summary judgment brief citing internal OpenAI/Microsoft messages calling AI training "astonishing theft" and admitting chatbots substitute for journalism18 Sept 202686
- US Commerce Department issues interim rule requiring licenses for AI chip and server exports to data centers that could provide compute to China and other restricted countries16 Nov 202684
- Anthropic launches Nasdaq IPO roadshow, pricing 825 million shares at $70-$78 for up to $90.4 billion total raise at up to $1.02 trillion valuation10 Nov 202680
- Anthropic files public S-1 with SEC for Nasdaq IPO under ticker "ANT"26 Oct 202680