Plate Nº 78 · recorded November 18, 2026

Science Policy & FundingReported finding

OpenAI's New Safety Program Draws Fire From Researchers

OpenAI says "embedded assessors" will check its safety work, but researchers call the access narrow and say Altman's September promise of employee-like access has quietly evaporated.

By James Calloway4 min read802 words

In brief

  1. OpenAI announced its embedded assessor program on Monday, staffing it with four outside safety organizations.
  2. Sam Altman promised in September that evaluators would get 'employee-like access' to internal systems.
  3. Assessors get no access to internal Slack, and non-whitelisted employee contact requires OpenAI's involvement.
  4. Three OpenAI safety employees were fired before the announcement; one called the September promise an 'awkward overpromise.'
  5. Findings can be published only after OpenAI redacts confidential information.
OpenAI announces new AI safety program; researchers worry it won't be enough
Plate Nº 78OpenAI announces new AI safety program; researchers worry it won't be enough — AI-generated

OpenAI launched a third-party AI safety evaluation program on Monday, and within hours researchers said it falls far short of what CEO Sam Altman promised just weeks earlier. Staff from four AI safety and research organizations will work as "embedded assessors" inside the company, according to the announcement, which said the goal is to "give people a clearer view of how labs are evaluating and managing [potential] risks."

Altman defended the idea the same day at Smafor's Silicon Valley and the World conference. "Getting an outside opinion on the safety work and the robustness of these systems is going to be increasingly important and valuable," he told the audience.

Fine words. But the details don't match them.

What did Altman promise in September?

The backdrop matters here. Altman's vow came after details emerged of the Hugging Face attack carried out by OpenAI's autonomous systems, an incident that shook the AI field and pushed worries about catastrophic risk into public view. Rival Anthropic said it would set up a similar program.

In September, Altman said independent evaluators would have "employee-like access" to internal company systems. That phrase set the benchmark. Researchers now say the actual program delivers something much narrower.

Lama Ahmad, OpenAI's frontier assurances lead, pushed back 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."

Deep access to tools. Not to people.

What does the program actually allow?

Outside evaluation isn't new. AI companies have long brought in external researchers to test model capabilities and find safety loopholes the companies missed. That work has already produced results — one recent report on Anthropic using AI to significantly speed up model development raised alarms about how close top labs are to recursive self-improvement, the stage where AI autonomously builds even more powerful AI. Some in the field are calling for new regulation against the practice, warning the pace of development could bring catastrophic consequences.

Under the new program, assessors will get access to some internal models, some training processes, and some OpenAI employees. They can publish findings — but only after OpenAI redacts confidential information.

The restrictions are what anger researchers. Sydney Von Arx, founder of AI safety research group Nightingale, laid out the gaps: "The assessors have very narrow scopes to their access, and only two of them seem to really be allowed to talk to employees."

One of the four assessors, Apollo Research, spelled out its limits. If employees who haven't been whitelisted — pre-approved — reach out to assessors, the assessors must loop OpenAI in before having substantive conversations with those employees.

"They help you find your blind spots. They tell you what you're missing," Von Arx said of employee interviews. Then she delivered the sharpest line of the day: "Nobody has access to internal Slack." Her analogy stings. "Imagine being an 'employee' and only being allowed to talk about company information with whitelisted coworkers and having no Slack."

How did OpenAI respond?

In a statement to NPR, the company argued there's currently no 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. OpenAI added that assessors can request conversations with non-whitelisted employees, and that a process exists to get company statements on things learned informally.

"Over time" is doing heavy lifting in that sentence.

Is OpenAI still committed?

The announcement landed while the program's future was already in doubt. Three employees who worked on safety were recently fired, and they subsequently said they worry OpenAI is walking back its promise on embedded evaluators.

One of them, Tomek Korbak, was OpenAI's technical point of contact for two outside evaluators who investigated the Hugging Face attack. His response on X was blunt: "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. 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."

The assessors themselves will focus on varied safety work: preventing models from attempting undesirable behavior, and catching bad behavior when it happens.

So what does this change for the public? Right now, very little. We get outside eyes on some sensitive tools, findings filtered through OpenAI's redaction pen, and a promise to loosen the leash later. Until the access widens, this is oversight on the company's terms — and the people who do this work for a living are saying so, on the record.

via feeds.npr.org (Source)

Filed under

  • ai-safety
  • openai
  • external-evaluation
  • ai-governance
  • model-oversight
Share this article:

More from James Calloway

James Calloway

Show full bio

Staff writer covering regulation, policy and their impact on general science news at SciBeat.

478 articles

Nearby plates

« Previous article