OpenAI's Embedded Safety Assessors Fall Short of Altman's Pledge
OpenAI embedded Apollo, METR, Transluce and Goodfire as safety assessors, but researchers say the terms fall short of Altman's September "employee-like access" pledge.
By Amara Osei
4 min read

What's News
- OpenAI signed agreements with four research organizations — Apollo Research, METR, Transluce and Goodfire — on Monday, November 2025.
- Researchers say the terms fall short of the "employee-like access" Sam Altman promised in September 2025.
- Transluce plans a follow-on eight-week assessment with broader access in early 2027.
- METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date.
- Three OpenAI safety employees were fired, including Tomek Korbak, who called Altman's pledge an "awkward overpromise."
OpenAI signed agreements with four independent research organizations on Monday to embed them as internal safety assessors — but the terms fall short of the "employee-like access" CEO Sam Altman promised in September, according to AI safety researchers who reviewed the program.
Apollo Research, METR, Transluce and Goodfire will operate as "embedded assessors" with what OpenAI calls "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." The company said the work will focus on alignment and safeguards during training and internal use, complementing its safety cases and supporting commitments under the White House Accord on Super Intelligence.
Altman defended the pilot at Semafor's Silicon Valley and the World conference Monday. "Getting an outside opinion on the safety work and the robustness of these systems is going to be increasingly important and valuable," he said.
What did Altman promise in September?
The gap dates to September, when details of the Hugging Face attack by OpenAI's autonomous systems shook the field. Researchers say Altman then vowed that independent evaluators would receive "employee-like access" to internal company systems. Rival Anthropic said it would set up a similar program.
Monday's announcement offers less. OpenAI said assessors get "model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible of the work." Assessors will work on-site using OpenAI devices, and the Safety and Security Committee of the OpenAI Foundation's Board will receive unredacted reports.
How narrow is the access in practice?
Sydney Von Arx, founder of Nightingale, said the restrictions don't resemble employment. "The assessors have very narrow scopes to their access, and only two of them seem to really be allowed to talk to employees," Von Arx said.
"Nobody has access to internal Slack," Von Arx added. "Imagine being an 'employee' and only being allowed to talk about company information with whitelisted coworkers and having no Slack."
Apollo Research's terms state that if non-whitelisted employees reach out to assessors, the assessors must loop OpenAI in before any substantive conversation. Conversations with authorized staff are off the record and cannot be published or attributed. A separate attestation pathway lets assessors request OpenAI-attested statements verifying or clarifying informal learnings — statements OpenAI may decline to provide.
What did the fired safety staff say?
Three OpenAI employees who worked on safety were fired. One of them, Tomek Korbak — who served as OpenAI's technical point of contact for two outside evaluators investigating the Hugging Face attack — posted on X that the program doesn't match the pledge.
"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," Korbak 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."
OpenAI spokesperson Gaby Raila 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," Raila said. The company added that assessors can request conversations with non-whitelisted employees and that a process exists for obtaining attested statements.
What will each assessor actually do?
The four engagements differ in scope:
- 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.
- Goodfire will embed interpretability researchers to analyze model internals during training runs, primarily for internal OpenAI researchers, while publishing learnings.
- METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date and conduct a broader monitor assessment pending scope agreement.
- Transluce will evaluate post-training alignment evaluations, with a follow-on eight-week assessment planned for early 2027 with broader access.
Who controls publication?
Assessors keep independence in publishing, but OpenAI retains the right to redact confidential information and may request responsible disclosure delays so it can remediate safety issues. Disagreements escalate first to the Safety Advisory Group and OpenAI executives, then to the Safety and Security Committee.
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.
The open question is whether the granted access lets assessors uncover failures OpenAI hasn't anticipated — the program's stated goal — or whether the limits on employee communication, Slack access and publication redactions will blunt exactly that capability.
Source: epyhia.com
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,260 articles