# OpenAI Fires Three Safety Researchers Over Alleged Leak to an Outside Safety Group

The dismissals land three days after OpenAI's president signed a White House accord promising internal verification teams and external audits of frontier models, with no enforcement attached.

- Published: 2026-10-02T06:31:19.109Z
- Canonical: https://polylog.news/ai/2026-10-02/openai-fires-three-safety-researchers-over-alleged-leak-to-a
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news)

OpenAI has terminated three members of its safety team after an internal investigation concluded they shared confidential company information with an outside AI-safety organization, [first reported by The Wall Street Journal](https://t.me/aipost/8332) and since confirmed in part by the company. An OpenAI spokesperson said the company "parted ways with three individuals for violating our policies on accessing and handling sensitive company information," and that the investigation found the individuals "mishandled sensitive information outside established company procedures," [according to Decrypt](https://decrypt.co/379857/openai-fires-researchers-alledged-leak-safety-group) and [Gizmodo](https://gizmodo.com/openai-ousts-three-safety-researchers-for-allegedly-mishandling-sensitive-information-2000820515).

What is verified is narrow. OpenAI has not named the three people, the organization that received the material, or what the material contained. Separate reporting noted the near-simultaneous departures of researchers Jasmine Wang, Mikita Balesni and Tomek Korbak, and neither OpenAI nor the researchers have said whether those exits are connected to the investigation.

The timing matters because of what OpenAI committed to days earlier. On September 29, OpenAI president Greg Brockman joined the chief executives of Google, Anthropic, Meta, Nvidia and xAI in [signing a voluntary safety accord with President Donald Trump](https://france24.com/en/technology/20260929-trump-says-ai-companies-sign-voluntary-accord-on-safety-controls), structured around four layers: internal controls, internal verification teams, external audits and independent board oversight. Trump called it morally binding. [No regulator enforces it](https://www.medianama.com/2026/09/223-trump-google-anthropic-meta-openai-ai-safety-accord/), and it carries no penalties and no deadlines.

## What this means

The mechanism at issue is information flow. An audit-and-verification regime that labs police themselves only works if evaluators outside the company can receive technical material, and OpenAI has just established that unapproved transfers of that material end employment. The exposed parties are third-party evaluators, red-team contractors and the safety staff who feed them, plus any enterprise buyer who treats the White House accord as a procurement assurance. The 2024 precedent is instructive: after the superalignment team broke up, researchers published a right-to-warn letter, policymakers held hearings, and no binding disclosure rule followed. The likelier outcome over the next several months is a formalized, narrow disclosure channel, with labs defining which outside bodies may see what and under which non-disclosure terms, rather than any expansion of independent access. A public filing by one of the dismissed researchers alleging retaliation, or a state attorney general opening an inquiry, would point the other way.

## What to watch

- Whether any of the dismissed researchers speaks publicly or files a whistleblower complaint, which would convert an internal personnel matter into a legal test of what safety staff may disclose.
- Whether the signatories of the September 29 accord name actual external auditors and specify their access to weights, evals and incident logs, the detail that separates a real audit from a press release.
- Whether other labs tighten or clarify their own rules on contact with outside safety organizations in response.