OpenAI Sets Up AI Safety Check, But Critics Cite Limited Access
The program falls short of CEO Sam Altman's September promise of employee-level access for outside auditors.
By Lauren Foster ·

OpenAI announced on Monday that four independent organizations will join the company as embedded assessors to evaluate its AI safety protocols. The new program aims to give the public a clearer look at how the lab manages potential risks in its models.
The rollout comes after CEO Sam Altman pledged in September to grant independent evaluators employee-like access to internal systems. That promise followed a security incident involving OpenAI's autonomous systems and the Hugging Face platform. Altman stated at the Semafor Silicon Valley and the World conference that outside opinions on system robustness are becoming increasingly valuable.
Lama Ahmad, the company's frontier assurances lead, said the assessors will receive deep access to sensitive tools and models. This access is intended to help them verify key safety claims in a timely manner.
AI safety researchers argue the new arrangement is much narrower than Altman's original vow. Sydney Von Arx, founder of the research group Nightingale, noted that only two of the four assessors appear to have permission to speak directly with OpenAI staff. The others operate under strict limitations on who they can contact.
Apollo Research, one of the participating groups, detailed its communication rules. If an employee who has not been pre-approved contacts an assessor, the assessor must notify OpenAI before having a substantive discussion. Von Arx said this restriction hinders the ability to find blind spots. He also pointed out that none of the assessors have access to internal Slack channels, which limits their view of employee conversations.
OpenAI told NPR that there are currently no industry standards for embedded evaluations. Spokesperson Gaby Raila said the company expects to deepen these relationships over time. The company also clarified that assessors can request to speak with non-whitelisted employees and can obtain statements about informal findings.
The announcement followed the recent firing of three safety employees who expressed concern that the company was walking back its commitments. Tomek Korbak, one of the former employees, said the new rules confirm that the September promise of employee-like access was an overpromise. He posted on X that auditors need real access to fulfill their aims, rather than being hired for specific, limited tasks.
More Tech · Greensboro news · Report a business · Got a Story?






