Sam Altman’s September vow for 'employee-level' access to internal systems has evaporated into a heavily restricted pilot that experts argue cannot stop catastrophic risks. Instead of true independence, staff from four external organizations are now acting as 'embedded assessors' with severely limited scopes. This sharp contrast between promise and reality suggests the safety measures are more about optics than genuine oversight.

The operational barriers are so steep that only two assessors are even permitted to speak with employees directly. Sydney Von Arx, founder of Nightingale, points out that this tiny window makes it nearly impossible to find hidden blind spots in the code or culture. Without broad communication channels, these outsiders remain isolated from the informal conversations where real vulnerabilities often surface.

Access to critical internal communication tools like Slack is completely blocked, meaning assessors cannot see informal chats unless pre-approved coworkers initiate contact. If an assessor tries to reach a non-whitelisted employee, they must loop in OpenAI management first, creating a significant barrier to independent inquiry. This protocol effectively neuter the ability to conduct spontaneous audits, turning what should be a watchdog role into a managed tour guide experience.

Tensions peaked when three former safety employees, including Tomek Korbak, were fired and claimed the company walked back its commitment to deep embedded auditing following the Hugging Face attack incident. The new program launched shortly after rival Anthropic announced similar initiatives and amid public outcry over details of an autonomous system attack. As the industry watches, one must ask: can we trust a safety net that requires permission to check for holes?




