The OpenAI logo is displayed on a smartphone screen placed on a reflective surface on which the company’s logo is projected in Creteil, France, on November 6, 2026.
Samuel Boivin/NurPhoto via Getty Images
hide title
Samuel Boivin/NurPhoto via Getty Images
AI safety researchers say OpenAI’s new external evaluation program falls far short of what the company promised, raising concerns about its effectiveness.
The company announced Monday that staff at four AI research and safety organizations will become “embedded testers” to “give people a clearer view of how labs test and manage.” [potential] risks.”
“Getting an outside opinion on the security work and robustness of these systems will be increasingly important and valuable,” CEO Sam Altman said at Semafor’s Silicon Valley and the World conference on Monday.
However, many AI security researchers say the new program is a far cry from the promise Altman made in September that independent testers would have “employee-like access” to the company’s internal systems.
Altman’s promise came after details of the Hugging Face attack by OpenAI autonomous systems shook the AI field. Concerns about the technology’s catastrophic risk captured public attention and rival Anthropic said it would establish such a program.
“We’re providing deep access to some of the most sensitive tools and models our researchers use… that are most important for evaluating key, relevant and timely security assertions,” Lama Ahmad, OpenAI’s frontier assurance lead, said in X.
Prior to these new integrated testing programs, AI companies hired outside researchers to evaluate the capabilities of AI models and identify security gaps that companies had overlooked.
The work of third-party evaluators has already helped shed light on the work of companies and increased public awareness of the security risks posed by the technology. A recent report on Anthropic’s use of AI to significantly speed up model development has raised concerns in the field about how close major companies are to using AI to autonomously build even more powerful models, a stage known as recursive self-improvement. Some in the field are concerned that if this happens, the speed of development could have catastrophic consequences for humanity and are demanding new regulation against the practice.
Researchers say the actual announcements about how these programs will work suggest something more limited than what Altman promised.
“The advisors have very limited scopes of access, and only two of them seem to actually have permission to talk to employees,” said Sydney Von Arx, founder of AI security research group Nightingale.
One of the announced testers, Apollo Research, described its limited access to OpenAI employees. If employees who have not been whitelisted or pre-approved communicate with the built-in testers, then the testers will need to connect OpenAI before having substantive conversations with the employees.
It’s important that outside investigators can interview employees, Von Arx said. “They help you find your blind spots. They tell you what you’re missing.”
Another glaring omission: “No one has access to internal Slack,” Von Arx said, referring to the corporate messaging tool, which would allow testers to see what employees are talking about with each other. “Imagine being an ’employee’ and only being able to discuss company information with whitelisted coworkers and not having Slack.”
In a statement to NPR, OpenAI said there are currently no standards for integrated assessments. “We look forward to deepening these relationships over time and sharing what we learn to help establish best practices for effective integrated assessments across the industry,” said OpenAI spokesperson Gaby Raila. The company also said that testers can request to speak with employees who have not yet been whitelisted, and that there is a process to obtain statements from OpenAI about things they have learned informally.
According to OpenAI, third-party organizations will focus on various aspects of AI security. Part of the work is related to preventing models from trying to engage in undesirable behavior, and other parts are related to detecting bad behavior when it happens. Testers will have access to some internal models, training processes, and some OpenAI employees. Testers will be able to publish their findings after OpenAI removes sensitive information, the company said.
OpenAI’s commitment to integrated assessments was questioned when three employees working in security were laid off. They later said they were concerned that OpenAI was backtracking on its promise to have built-in testers.
One of the fired employees, Tomek Korbak, was OpenAI’s technical point of contact for two external testers who investigated the Hugging Face attack.
“Integrated auditing will only meet its objectives if auditors actually gain access at the employee level rather than being hired to perform a specific job,” he posted on X in response to OpenAI’s announcement. “Unfortunately, OpenAI now makes it clear that [Sam Altman’s] “The September tweet about ’employee-like access’ was simply an awkward over-promise that the company wants to quietly back off.”