1. Home
2. News
3. OpenAI

Ryan Greenblatt calls OpenAI's leak accusations false, backing three fired safety researchers

Greenblatt confirms disappearing-message Signal chats with all three, but says nothing confidential went over them.

By Think Facility · October 23, 2026 · OpenAI AI safety Save

On October 23, 2026, Ryan Greenblatt wrote on X that OpenAI had made "false accusations" in the case of the three safety researchers it fired. According to the post, the three were told Greenblatt helped leak details of OpenAI's model architecture to The Information. Greenblatt's answer: "This is false."

The three are Tomek Korbak, Jasmine Wang and Mikita Balesni. They'd already denied being the source, in a letter to OpenAI's safety committees. Now the outsider they were linked to backs them up.

Accused

Ryan Greenblatt

Greenblatt's answer

This is false

Signal chats

disappearing messages on

Wang's comparison

nuclear and finance

What Greenblatt says

There are two denials. The first is simple. A reporter from The Information did get in touch shortly before the story ran, Greenblatt writes, and got no comment at all (just a request to make clear that a blog post the article cited was already public).

The second is messier. It concerns an internal OpenAI proposal on monitorability. Apparently one version of it held highly sensitive architecture details, Greenblatt has heard, and OpenAI accused Greenblatt of reading that version. Wrong document, according to the post. The one Greenblatt saw had no sensitive IP and was approved for sharing.

Greenblatt thinks OpenAI now knows it conflated the two. Whether the company has clearly taken the accusation back is another matter, and Greenblatt doesn't know.

Then there's Signal. Yes, Greenblatt had chats with all three with disappearing messages on. They're longtime professional contacts, the post says (and Signal is common in tech). During the Hugging Face investigation Greenblatt kept confidential information out of Signal messages, though high-level things sometimes came up on calls.

What the fired researchers say

After they were fired, the three wrote to OpenAI's Safety and Security Committee, its Safety Advisory Group and its Mission Advisory Council. They weren't the source, the letter says. They don't know who was. And the article actually cut against their own push for cross-company limits on unmonitorable architectures, so they'd no reason to leak it.

Two of them had jobs that meant talking to outsiders. Korbak was METR's technical point of contact in the Hugging Face incident investigation (METR is one of the outside evaluators OpenAI works with). Balesni was working across companies on commitments to prevent loss of monitorability, with board members and the C-suite in the loop. Wang's case is different: access to an executive's email, delegated for recruiting, that she'd asked IT to remove.

Their first request is that OpenAI keep its public commitment to embed outside safety auditors. They point to Sam Altman's September 12 promise to give independent evaluators employee-like access, made the day Dario Amodei published his essay on pacing AI (our story).

Late on October 22, Korbak offered the best explanation the three have so far.

How others read it

Support came quickly from people who work on AI risk. Nate Soares fixed on the timing, and on what the three were never told.

Granted, this is one side. OpenAI's account reaches us only through press reports (so it isn't quoted here), and Greenblatt's post doesn't say whether OpenAI has replied.

Who protects people who talk to auditors?

Wang made the wider argument on X the same night. In nuclear power and finance, Wang wrote, employees have defined channels for raising safety concerns confidentially with auditors and regulators, and AI needs rules that clear. An outside evaluator means little, in Wang's view, if staff can't speak candidly with it.

The nuclear version exists. The Nuclear Regulatory Commission's Allegation Program takes concerns about its requirements from licensees, their contractors and vendors, and the employees of any of them. Retaliation is handled elsewhere: discrimination claims under the Energy Reorganization Act go to the Department of Labor.

California has something narrower for AI. Senate Bill 53 (signed by Governor Gavin Newsom on September 29, 2025) protects whistleblowers who disclose significant health and safety risks posed by frontier models. It's about disclosing risks, though. The sources here don't settle whether ordinary work with an outside evaluator counts, and I think that's the gap Wang is pointing at. Neither the letter nor the posts mention the law.

The letter ends on a practical ask. OpenAI should set out how employees may work with outside safety organizations, so nobody has to guess where the lines are now, and share the letter widely inside the company.