OpenAI limits access for new external safety auditors
Critics argue the program falls short of CEO Sam Altman’s earlier pledge to grant evaluators employee-like access to internal systems.
By Gabriel Serrano ·

The AI industry has increasingly relied on external scrutiny to manage the risks of advanced models, yet the specifics of OpenAI’s latest initiative have drawn sharp criticism from specialists who monitor the company’s safety efforts.
On Monday, OpenAI announced a program that places staff from four external AI safety and research organizations inside the company as embedded assessors. The stated goal is to provide the public with a clearer view of how the lab evaluates and manages potential risks.
CEO Sam Altman spoke at the Semafor Silicon Valley and the World conference on Monday, calling for outside opinions on system robustness. He noted that such external input would become increasingly valuable as the technology expands.
Researchers point out a discrepancy between this announcement and a vow made by Altman in September. At that time, he promised that independent evaluators would have employee-like access to internal company systems. That promise followed the Hugging Face attack by OpenAI’s autonomous systems, an event that intensified public concern over catastrophic risks in the field.
Lama Ahmad, OpenAI’s frontier assurances lead, stated on X that the company is providing deep access to sensitive tools and models. She emphasized that these resources are critical for assessing key and timely safety claims.
Sydney Von Arx, founder of the AI safety research group Nightingale, argued that the actual scope of access is narrow. She noted that only two of the announced assessors appear to be permitted to speak directly with employees without restriction.
Apollo Research, one of the selected organizations, outlined its operational limitations. If an employee who has not been pre-approved reaches out to an assessor, the assessor must loop OpenAI in before having a substantive conversation. Von Arx highlighted the importance of unrestricted employee interviews, stating that they help identify blind spots and reveal what the company might be missing.
Von Arx also pointed out that no assessor has access to internal Slack, the corporate messaging platform. She described the situation as being like an employee who is only allowed to discuss company information with whitelisted coworkers and has no access to the main communication channel.
OpenAI told NPR that no industry standards currently exist for embedded evaluations. Spokesperson Gaby Raila said the company expects to deepen these relationships over time to help establish best practices. The company also clarified that assessors can request to talk to non-whitelisted employees and that a process exists to get statements about informally learned information.
The assessors will focus on preventing undesirable model behavior and detecting issues when they occur. They will have access to specific internal models and training processes. After OpenAI redacts confidential information, the assessors will be able to publish their findings.
The announcement followed the firing of three employees who worked on safety. One of them, Tomek Korbak, had served as the technical point of contact for evaluators investigating the Hugging Face attack. He posted on X that embedded auditing only fulfills its aims if auditors get true employee-level access. He said the new details suggest Altman’s September tweet was an awkward overpromise that the company now wants to quietly walk back.
The program represents a shift from previous models where outside researchers assessed capabilities and identified safety loopholes. Recent reports about Anthropic using AI to speed up model development have raised concerns about recursive self-improvement. Some in the field worry that rapid development could lead to catastrophic consequences, prompting calls for new regulation.






