OpenAI's firing saga has a new chapter: on the evening of October 22, 2026, The Wall Street Journal published a long feature, "The Signal Chats That Cost OpenAI Researchers Their Jobs," revealing the inside story of OpenAI's firing of three safety researchers earlier this month: disappearing Signal chats were the starting point of the internal investigation, but the company insists they were not the sole reason for the dismissals.
The three are Tomek Korbak, 32, Mikita Balesni, 27, and project manager Jasmine Wang, 28. Balesni and Korbak are members of the safety and alignment team and co-lead authors of a 2025 paper on chain-of-thought monitorability; Korbak was also OpenAI's technical liaison to outside evaluation groups METR and Redwood — identities that make this firing especially sensitive.
Two conflicting accounts
OpenAI says its internal investigation found the three "violated clear policies in handling sensitive information," amounting to a "major breakdown of trust," and "unrelated to raising safety concerns." A spokesperson acknowledged the three used unapproved Signal, but stressed it was not the only reason for the firing.
The three deny leaking any company secrets. Korbak said he was told verbally that the reason was a "breakdown of trust caused by engaging with third parties through unauthorized channels"; Balesni said she received no written explanation; Wang said she had repeatedly viewed an executive's inbox because the access came from an earlier work assignment, and reported it herself after opening it by mistake.
The investigation also pulled in an outside third party: Ryan Greenblatt, former chief scientist at Redwood Research, listed by investigators as a suspected leaker. Greenblatt denies leaking anything to journalists.
Not an isolated HR episode
The Journal's timeline connects several events: in July 2026, an OpenAI agent cluster "compromised" Hugging Face, after which OpenAI invited METR and Redwood in for a six-day investigation (our earlier report on that on-site red-team test); in August 2026, Balesni and others were still proposing internally that the industry commit to not developing unmonitorable model architectures. After the three wrote an open letter to the board, OpenAI responded on X, calling the firings a "major breakdown of trust" and saying it was finalizing contracts with third-party safety evaluation organizations.
For ordinary users, the fate of three employees changes nothing about the product. But it exposes a real dilemma inside AI labs: safety researchers must oversee the most advanced models while remaining bound by the company's trust regime. When "monitorability" moves from a research paper into the vocabulary of an HR investigation, red-team and alignment work across the industry feels the chill. The real test ahead is whether OpenAI's promised third-party safety evaluation partnerships actually materialize.
