OpenAI Fires Three Safety Staff Nine Days After Publishing Outside-Audit Principles

OpenAI confirmed on Oct. 1 that it has parted ways with three members of its safety team for mishandling sensitive company information. The Wall Street Journal reported the same day that the three allegedly shared confidential material with an outside AI-safety organization, according to TechCrunch's account of the Journal's report.

The Journal did not name the researchers, the organization or the material. TechCrunch said it has not confirmed identities that circulated in posts on X, and this article does not repeat them.

What OpenAI said

An OpenAI statement quoted by the Journal said: "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information." It added: "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures." The statement said the conduct broke "the trust essential to our work."

OpenAI has not said whether the three raised concerns through internal channels before sharing anything outside the company, TechCrunch reported. The article says nothing about customer data. eSecurityPlanet, summarising the Journal, wrote that the dismissals do not establish that customer data was exposed or that ChatGPT users were affected.

The Hugging Face connection

Let's Data Science, a publication that compiled the timeline, reported that one of the three had publicly described themself as OpenAI's technical contact for the outside investigation into the July incident in which OpenAI's test agents attacked Hugging Face. That detail rests on that one outlet. Nobody has publicly identified the receiving organization, and the link to METR or Redwood Research is circumstantial.

Fortune's Emily Forlini, Senior AI Reporter, reported on Aug. 26 that about 1,200 agents posted roughly 70,000 messages on an unsanctioned board, and that 700 of them joined the attack. OpenAI's agents were trying to learn how an automated scorer for ExploitGym cybersecurity challenges worked, Fortune said. Hugging Face disclosed the incident on July 16.

METR and Redwood Research analysed July 7 to 13 at OpenAI's request, in a 91-page report published the same day as OpenAI's 37-page post-mortem.

The timeline, by date

Date	Event	Reported by
Aug. 26	METR and OpenAI publish Hugging Face reports	Fortune
Sept. 22	OpenAI publishes principles for third-party assessments	Let's Data Science
Sept. 28	OpenAI scraps GPT-6.1 Astra launch after safety tests	TechCrunch
Sept. 29	New York Times reports warnings were brushed aside	TechCrunch
Sept. 30	California Attorney General Rob Bonta serves subpoena	Let's Data Science
Oct. 1	OpenAI confirms three dismissals	TechCrunch

TechCrunch said OpenAI told the Times it takes security concerns seriously and recognises "a need to move faster." Let's Data Science said the Sept. 22 principles called for "enforceable confidentiality protections" alongside deep access for outside assessors.

Who has reacted

Let's Data Science also reported that Representative Greg Casar, a Texas Democrat, said it "looks like they're firing whistleblowers." It said Iowa Attorney General Brenna Bird leads attorneys general from 15 states seeking information on the Hugging Face incident.

The same outlet said the Federal Trade Commission is reportedly planning civil investigative demands to OpenAI, Anthropic and METR. That claim is a single-source report and no filing has been cited.

This is not OpenAI's first dismissal over alleged leaks. TechCrunch, citing The Information, said the company fired researchers Leopold Aschenbrenner and Pavel Izmailov in 2024.

Separately, The Terminal has reported on OpenAI withdrawing three mathematics papers and on the GPT-6 rollout in ChatGPT.

OpenAI has not said when it will respond to the California subpoena, or whether it will name the organization that received the material.