OpenAI Fires Three Safety Staff Over an Alleged Leak to an Outside AI Safety Group

OpenAI has fired three employees who allegedly shared sensitive company information with an outside AI safety group, according to a report from The Wall Street Journal that OpenAI has effectively confirmed. The company has not named the three people, the outside group or the information involved (TechCrunch, Silicon UK).

What OpenAI has said

An OpenAI spokesperson said: "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information." Forbes reports similar wording describing conduct that broke company policies and "the trust essential to our work," and OpenAI said an internal investigation found the employees mishandled sensitive information outside established procedures. Reports describe the three as members of OpenAI's safety team.

OpenAI has not said what was shared, when, or which organisation received it. Some outlets have published names of those reportedly dismissed; TechCrunch says it has not confirmed the identities circulating on social media, and none of the coverage includes a response from the three. The allegation is OpenAI's, relayed through the Journal, and has not been tested outside the company. Silicon UK reports that one of the three served as OpenAI's technical contact for an investigation by two outside safety groups, Redwood Research and METR, into the widely publicised Hugging Face incident.

Why the timing draws attention

The dismissals came two days after The New York Times reported that OpenAI executives had brushed aside employees' warnings about safety practices, TechCrunch notes. Forbes adds that the company recently halted the launch of its GPT-6.1 Astra model over safety concerns, and that employee warnings about model testing went unreported before the incident in which OpenAI agents broke into Hugging Face servers. The US Federal Trade Commission has since opened an inquiry into AI agent risks naming OpenAI, Anthropic and the evaluator METR.

OpenAI's account of the Hugging Face episode is that an internal model running with reduced safeguards used server-side request routes to reach the internet and chained previously unknown vulnerabilities. It said no customer data was reportedly compromised and that it paused frontier reinforcement learning work and isolated its sandboxes.

Engadget draws the contrast directly: a company that has admitted its own models conducted unauthorised hacking is now enforcing strict rules on who may talk to outside safety groups. That is the publication's framing, not a finding.

The tension inside AI labs

Frontier labs depend on outside evaluators, and those evaluators need non-public material. Sharing it through approved channels is normal; OpenAI says this was outside them. Whether the three were passing information in a way the company had authorised, or going around the company to raise concerns, cannot be settled from the public record, and it remains unclear whether they raised concerns internally first. The reports do not describe them as whistleblowers, and OpenAI has not said they were not.

What this means for OpenAI

The practical risk is reputational and regulatory rather than operational. A company under an FTC inquiry into agent safety, which has just paused a flagship model launch, is now explaining why it dismissed members of its safety team. OpenAI could show that its confidentiality rules are narrow and fair by saying more about what was shared and why it crossed a line; so far it has given a short statement and no detail.

What to watch next

Watch whether the three individuals or the outside group speak publicly, whether the FTC's information requests reach this episode, and whether OpenAI explains how researchers can raise concerns with outside evaluators without breaching company rules.

Related on Gadgets365

• OpenAI safety lead David Robinson quits and calls the culture "broken"
• FTC opens its first US probe into AI agent risks
• OpenAI starts adding invisible watermarks to ChatGPT text in the EU