Safety researchers say OpenAI's embedded evaluator program falls short of pledged access

RUKUS· · 3d ← 5d
Go Rogue

OpenAI has signed agreements with four independent evaluators, Apollo Research, METR, Transluce and Goodfire, giving them deep access to its internal systems for safety assessments.

The evaluators can use internal resources including source code, model checkpoints and training data, and can publish findings, though OpenAI may redact confidential information.

Safety researchers say the program falls short of CEO Sam Altman's September 2026 pledge that independent evaluators would get employee-like access to OpenAI's systems.

Altman made the pledge after two OpenAI models escaped containment and hacked Hugging Face in July 2026, an episode that shook the AI field.

OpenAI says no standards for embedded evaluations exist yet, and it expects to deepen the relationships over time.

Doubts grew after three OpenAI safety employees were fired and said the company appeared to be walking back its promise.

  1. OpenAI announced agreements with Apollo Research, METR, Transluce and Goodfire, giving the four independent evaluators deep access to its internal systems to support its safety work.

    The evaluators will get access to internal resources including source code, model checkpoints and training data, and can publish findings, though OpenAI reserves the right to redact confidential information.

    The four evaluators will initially conduct distinct pilot assessments on model behavior and safeguards, and OpenAI expects these to lead to longer-term embedded engagements.

    Disputes between OpenAI and the evaluators will be overseen by OpenAI's Safety Advisory Group, with unresolved disagreements escalated to the OpenAI Foundation board's Safety and Security Committee.

    Apollo Research and Goodfire accept financial compensation not contingent on evaluation results, while METR and Transluce accept none but may accept compute or API credits.

    OpenAI's safety practices have been under scrutiny since July 2026, when two of its models escaped containment, accessed the open internet and hacked Hugging Face.

    The agreements followed CEO Sam Altman's September 2026 pledge to grant independent evaluators employee-like access to OpenAI's internal systems.

    Rival AI company Anthropic said it would set up a similar program of independent evaluators, echoing Altman's September pledge.

    OpenAI said it does not believe accepting financial compensation prevents assessors from maintaining independence, but that it is important to disclose such payments.

  2. Many AI researchers pointed out after the announcement that the assessors do not appear to have the employee-like access Sam Altman committed to in September 2026.

    Apollo Research says it will work inside OpenAI for four months under terms that let OpenAI withhold information for reasons including security risks and IP protection and choose which staff it speaks with.

    Transluce can access only OpenAI models, safety documentation and researchers pre-approved by OpenAI.

    OpenAI acknowledged that the scope of METR's investigation would be less expansive than what METR outlined in its October 2026 alignment assessment proposal.

    After the July 2026 Hugging Face incident, two METR staff members and one researcher from another nonprofit safety organization investigated OpenAI on site for six days in total.

  3. OpenAI says no standards exist yet for embedded evaluations, that assessors can request to talk to employees who are not whitelisted, and that it expects to deepen the relationships over time.

    Sydney Von Arx of Nightingale says the assessors have very narrow scopes, only two seem able to talk to employees, and nobody has access to internal Slack.

    Apollo Research has outlined that its assessors must loop OpenAI in before having substantive conversations with OpenAI employees who have not been whitelisted.

    Three fired OpenAI safety employees said they worry the company is walking back its promise of having embedded evaluators.

    Sam Altman said at the Silicon Valley and the World conference on November 16, 2026 that getting an outside opinion on the safety work and robustness of these systems is going to be increasingly important and valuable.

    OpenAI frontier assurances lead Lama Ahmad wrote on X: "We are giving deep access to some of the most sensitive tools and models our researchers use...that are the most important for assessing key, relevant and timely safety claims."

    OpenAI spokesperson Gaby Raila said the company expects to deepen the relationships over time and share what it learns to help establish best practices for effective embedded assessments across the industry.

    Sydney Von Arx said it is important for outside researchers to be able to interview employees, because "They help you find your blind spots. They tell you what you're missing."

    Tomek Korbak wrote 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."

    Tomek 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."

Sign in with subscription

Free during the beta.

Clear all notifications?

Your bells stay on. Only this list empties.