Three former OpenAI safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—have publicly disputed the grounds for their recent termination. Fired following an internal investigation into alleged policy violations regarding sensitive information, the researchers told the Wall Street Journal that OpenAI never provided written documentation detailing what specific data they allegedly leaked. They deny sharing confidential architecture details and suggest their dismissal relates to internal advocacy for transparent AI reasoning models.

The terminations occurred after security investigators discovered unapproved Signal communications with disappearing messages on the researchers’ personal devices during a leak probe. The probe followed media reports detailing OpenAI’s Astra model. The researchers claim their external discussions focused on maintaining public and auditor access to internal model chain-of-thought traces, an issue central to AI safety monitoring.

Why it matters

  • Underscores growing tension between frontier AI lab confidentiality and external safety audit transparency.

  • Exposes operational risks and strict corporate surveillance surrounding model architecture intellectual property.

  • May deter independent safety auditors and researchers from collaborating closely with top AI developers.

Source: the-decoder.com