OpenAI Launches Embedded Safety Assessors, but Researchers Say Access Falls Short of "Employee-Like" Promise
OpenAI announced Monday a pilot program embedding four independent research organizations inside its operations to evaluate AI safety, but the access granted appears narrower than the "employee-like access" CEO Sam Altman pledged in September.
The company said it has signed agreements with Apollo Research, METR, Transluce, and Goodfire to serve as "embedded assessors" with "deep access to a lab's internal models and systems to assess model alignment and the effectiveness of safeguards over an extended period of time" [2]. The assessments will focus on alignment and safeguards during training and internal use, complementing OpenAI's safety cases and supporting commitments under the White House Accord on Super Intelligence [2].
At Semafor's Silicon Valley and the World conference Monday, Altman said, "Getting an outside opinion on the safety work and the robustness of these systems is going to be increasingly important and valuable" [1].
The September promise vs. the November program
In September, after details of the Hugging Face attack by OpenAI's autonomous systems shook the field, researchers say Altman vowed that independent evaluators would receive "employee-like access" to internal company systems [1]. Rival Anthropic said it would set up a similar program [1].
OpenAI's announcement describes providing assessors with "model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work" [2]. Assessors will work on-site using OpenAI devices, and the Safety and Security Committee of the OpenAI Foundation's Board will receive unredacted reports [2].
Researchers identify gaps
AI safety researchers say the practical terms fall well short of employee-level access.
"The assessors have very narrow scopes to their access, and only two of them seem to really be allowed to talk to employees," said Sydney Von Arx, founder of Nightingale [1]. Apollo Research outlined that if non-whitelisted employees reach out to assessors, the assessors must loop OpenAI in before substantive conversations [1].
"Nobody has access to internal Slack," Von Arx said. "Imagine being an 'employee' and only being allowed to talk about company information with whitelisted coworkers and having no Slack" [1].
Conversations with authorized staff are off the record and cannot be published or attributed. A separate attestation pathway allows assessors to request OpenAI-attested statements verifying or clarifying informal learnings, which OpenAI may decline to provide [2].
Fired safety staff call it a walk-back
Three OpenAI employees who worked on safety were fired. One, Tomek Korbak—who served as OpenAI's technical point of contact for two outside evaluators investigating the Hugging Face attack—posted on X: "Embedded auditing will only fulfill its aims if the auditors really get employee-level access as opposed to being hired to do a specific job," he wrote. "Sadly, OpenAI makes it clear now that [Sam Altman's] September tweet about 'employee-like access' was just an awkward overpromise that the company wants to quietly walk back" [1].
OpenAI told NPR there are currently no industry standards for embedded evaluations. "We expect to deepen these relationships over time and share what we learn to help establish best practices for effective embedded assessments across the industry," spokesperson Gaby Raila said [1]. The company added that assessors can request conversations with non-whitelisted employees and that a process exists for obtaining attested statements [1].
Scope varies by assessor
The four engagements differ in focus. Apollo Research will test whether models could covertly subvert developer intentions and whether safeguards would detect such behavior, in two stages across selected public and unreleased models [2]. Goodfire will embed interpretability researchers to analyze model internals during training runs, primarily for internal OpenAI researchers, while publishing learnings [2]. METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date and conduct a broader monitor assessment pending scope agreement [2]. Transluce will evaluate post-training alignment evaluations, with intent for a follow-on eight-week assessment in early 2027 with broader access [2].
Publication and oversight
Assessors maintain independence in publishing but OpenAI retains the right to redact confidential information and may request responsible disclosure delays to remediate safety issues [2]. Disagreements escalate to the Safety Advisory Group and OpenAI executives, then to the Safety and Security Committee [2].
OpenAI said it expects this first phase to lead to longer-term embedded engagements and is exploring future assessments of compliance with safety frameworks and organizational culture [2].
What remains unclear is whether the access granted will enable assessors to uncover failures OpenAI has not anticipated—the stated goal of the program—or whether the restrictions on employee communication, Slack access, and publication redactions will limit that possibility.