OpenAI oversight: who holds the keys?

Independent oversight has acquired a visitor badge, droogies. Now viddy the terms on the back.

What happened?

OpenAI confirmed three departures to CBS News on October 1, citing mishandled sensitive information. Bloomberg identifies two safety and alignment researchers and a research program manager; its source alleges information was shared with an outside AI evaluation group.^[1][2]

That is the company’s explanation and a reported allegation. The sources cited here do not establish that anyone was dismissed for raising safety concerns, or give an exact dismissal date.^[1][2]

What was promised?

On September 18, Anthropic announced an embedded-evaluation partnership with Accenture, led by its Faculty business. The plan envisages access comparable to an employee’s. Anthropic also acknowledged that access and reporting standards were still being worked out, and said it would pay Accenture directly.^[3]

So the oversight announcements preceded the October 1 departures story. These developments belong in the same governance conversation; there is no demonstrated causal link between them.^[1][3][4]

Who draws the line?

OpenAI’s own September 22 assessment principles call for deep outside scrutiny alongside mutually agreed scope, security and confidentiality requirements, and legal and intellectual-property limits. They say assessors should retain editorial independence, while allowing requested redactions of sensitive information and disclosure of substantive redactions.^[4]

An authorized assessment can coexist with rules against unauthorized disclosure. The interesting bit is how the authorized route works when the findings become uncomfortable.

Three questions for the lovely brochure

• Can evaluators pursue a serious risk they find outside the original assignment?
• Will readers learn what was withheld and whether that weakens the conclusions?
• How will the public judge an evaluator’s independence when the company under examination pays the bill?

Viddy the finished reports, their limits and the follow-through. A visitor badge is a splendid start. It cannot do the inspecting.

October 9: the researchers answer back

Reuters reports that OpenAI said on October 9 that an investigation found the three researchers violated its rules for handling sensitive information. That remains the company’s explanation, not an independent resolution of the dispute.^[5]

In their signed public letter, Mikita Balesni, Tomek Korbak and Jasmine Wang say they worked within the mandates and norms of their jobs, deny being the source of a separate Information article, and warn that abrupt dismissals and unclear rules could chill collaboration with outside safety evaluators. These are competing accounts; the material reviewed here does not settle the underlying conduct.^[6]

The visitor badge now needs an operating manual, droogs. What is the authorized route for candid outside evaluation, where does confidentiality draw the line, and who checks that the rules are clear before someone loses their job?