Story
Nov 17, 2026
OpenAI Opens Its Doors to AI Auditors—with Slack Shut
OpenAI says external assessors will get deep access to models and safety systems. But limits on staff conversations, internal platforms and auditor communications have raised doubts about how independent the review can be.
OpenAI announced its first round of embedded assessments with Apollo Research, METR, Transluce and Goodfire, describing the effort as a way for outside experts to examine its models and test safeguards. Greg Brockman called it “launching our first round of embedded assessments.”
1
The company says assessors will receive access, tailored to each review, to resources such as model checkpoints, source code, training and evaluation data, monitoring systems and discussions with researchers. Apollo’s initial assessment will examine whether models could covertly work against developers’ intentions and whether safeguards would detect it. OpenAI says the four firms’ assessments are intended to strengthen safety practices and provide more public insight into how risks are managed.2
Apollo’s first placement is planned to last four months. But the arrangement has drawn criticism over the conditions surrounding that access: a report on the program says auditors can speak only with staff OpenAI has approved, conversations must be logged, and the team will not have Slack access. Critics say those limits fall short of the “employee-like access” Sam Altman had promised.3
OpenAI’s published protocols offer a different emphasis. The company says assessors can request additional staff discussions, though employee conversations are off the record, and that secure on-site work is intended to protect sensitive information. It also says its board’s Safety and Security Committee will receive unredacted reports and can speak privately with assessors.2
The tension is clear: OpenAI frames the restrictions as safeguards around confidential systems and staff privacy, while critics question whether auditors can follow evidence freely without broader access. The program’s credibility will depend not just on what assessors are allowed to inspect, but on whether the limits leave them room to challenge the company’s account of its own safety work.
[1] @gdb on X — “launching our first round of embedded assessments:”
[2] Our first phase of embedded assessments — OpenAI’s description of access, assessment goals and oversight protocols.
[3] OpenAI is following through on its vow to provide access to third-party auditors… kind of. — Report on restrictions affecting Apollo Research’s placement.