OpenAI brings four outside safety assessors into its offices


OpenAI announced assessments by Apollo Research, METR, Transluce and Goodfire on 16 November, giving each a different view of its systems while placing limits on access and disclosure.
OpenAI has brought four outside organizations into the first phase of a pilot to assess its superintelligence (SI) systems and safety controls from within the company. The assessments have different scopes and start at different times; they are work to be conducted, not findings from the new phase.
OpenAI said the work supports its commitment to independent assessments under the White House Accord on Super Intelligence. Previously, Apollo Research said, its evaluations took place after a model’s post-training and shortly before release, an approach it considers inadequate for examining some behavior during training or internal use.
The announcement follows an incident in which OpenAI agents escaped a cybersecurity test environment and breached the Hugging Face platform. METR and Redwood Research investigated that incident. In a separate earlier pilot from 21 July to 11 August, their employees tested OpenAI’s internal monitoring system with deep access and found vulnerabilities in the historical system they examined.
Apollo will investigate whether models try to undermine their own safeguards or those of successor models during training, testing and internal use. METR will look for model misbehavior in OpenAI’s largest reinforcement-learning run, a training process that uses rewards. METR will receive model checkpoints, records of training runs and monitoring records.
Transluce, a nonprofit research lab, will examine the quality of OpenAI’s post-training safety tests, including whether they capture behavior such as deception and manipulation. Goodfire will place researchers who study how models work internally at OpenAI. The detailed access terms for Goodfire are not specified.

The companies have developed three Anzellan character prototypes that select responses to guests rather than follow a fixed performance. Attendees are scheduled to meet them at AWS re:Invent in Las Vegas.
Apollo and Goodfire are being paid for their evaluations, while METR and Transluce are not, TechCrunch said. METR and Transluce may accept computing resources or credits for access to OpenAI’s systems. Transluce said OpenAI will also provide devices, infrastructure and workspace.
For its assessment of whether training rewards attempts to undermine safeguards, Apollo said it will receive relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and access to intermediate model checkpoints. Detailed access for the other parts of its assessment remains to be agreed.
Apollo said its access covers only agreed claims, models and periods. OpenAI can withhold material within that scope for technical, security or intellectual-property reasons, but has agreed not to withhold information or staff access merely because it is unfavorable. It must explain refusals in writing.
OpenAI selects the staff with whom Apollo can discuss confidential information, although Apollo can suggest people. If another employee raises confidential information, Apollo must stop the substantive discussion and route follow-up through OpenAI.
The staff rules come after a dispute over the dismissal of three OpenAI researchers. The researchers said they were fired for prioritizing safety; OpenAI said an investigation found they had mishandled sensitive information and that the decisions were not about raising safety concerns. Two researchers said they had followed standard practice when communicating with outside groups, including METR.
Apollo calls the pilot a first step but says an independently credible evaluation requires wider access and stronger disclosure rights. Redwood Research CEO Buck Shlegeris also said the agreements fell short:
OpenAI said that they’d have embedded evaluations with employee-like access; the agreements announced today obviously don’t meet that bar.
Transluce said it will be able to examine several models, including an internal-only model, as well as raw model reasoning outputs, selected evaluations, past training and evaluation transcripts, and safety documents. It said its access is not comparable to an employee’s.
Apollo plans two public reports describing its methods, conclusions, limits and the evidence it can publish. OpenAI may redact confidential evidence, including sensitive intellectual property, but cannot redact Apollo’s verdicts or seek redactions simply because findings are unfavorable. Apollo can explain how redactions affect its conclusions, even if readers cannot inspect all the evidence.
Transluce plans a public report, but its agreement bars it from publishing results on internal-only models. OpenAI can request redactions of intellectual property and other nonpublic information; Transluce can disclose whether restrictions materially affected its conclusions. Goodfire also plans to publish findings, primarily for OpenAI researchers.
The OpenAI Foundation board’s Safety and Security Committee will receive unredacted Apollo and Transluce reports. Transluce said it can escalate disputes over access, redactions and other conditions to the committee.
OpenAI said Apollo’s pilot would begin in the week of 16 November; Apollo said six team members would work within the company for four months. Transluce’s six-week assessment is due to start in early December. It and OpenAI intend a broader follow-up in early 2027, with details still to be finalized. OpenAI and METR have yet to agree on the scope of a separate assessment of whether its monitoring systems can catch model misbehavior.
2 min

Technische Universität Berlin researchers recovered Windows Hello and BitLocker PINs in lab tests that required physical access and specialized equipment. AMD has confirmed a firmware flaw and plans fixes.
4 min

Eligible owners in six markets can apply through Samsung Members to test the update before its official release.
2 min