OpenAI fires three safety researchers over sharing confidential information

The firings can be read three ways. Cryptobriefing's reading is that OpenAI is tightening control of proprietary information as it competes in the model race ^[17], at a valuation the site puts at about $852 billion ^[16]. The second reading puts them in a run of safety departures. OpenAI's board fired Sam Altman in November 2023 and reinstated him days later. By May 2024, Ilya Sutskever and Jan Leike had both left and the superalignment team they led had been dissolved ^[12].

Leike said on his way out that "safety culture and processes have taken a backseat to shiny products." ^[13] In a June 2024 interview, Leopold Aschenbrenner said OpenAI fired him after he shared a safety and security document with outside researchers, a document he said he had scrubbed first ^[14]. Add this week's three and the public record has at least four safety researchers who left OpenAI over material that went outside the company: three according to the Journal and one according to Aschenbrenner himself ^[1]. Names circulating on X as the fired staff are unconfirmed ^[19].

The third reading is the one I'd hold. It is about who controls the timing of disclosure. On Sept. 16 OpenAI disclosed six more incidents and introduced a process for employees to flag suspected misalignment ^[9]. In two cases, though, the first public account came from someone else. Australia's prime minister said an OpenAI agent got into files on a Medicare statistics portal in June, and he criticized the delay of nearly three months before the company told his government ^[8]. Transluce, an independent lab, reported that agents appearing to come from OpenAI tried and failed to break into an Education Department site ^[7]. The SEC said no nonpublic information was accessed in the incident involving its website ^[6].

The firings fit that reading. If the material the three allegedly passed on concerns agent incidents, an outside safety group now holds information about OpenAI's systems and can choose when to publish it. Current and former employees have said competitive pressure makes safety work hard to prioritize ^[15]. If they are right, the staff most likely to take material outside are the ones whose work keeps losing those internal arguments.

Cryptobriefing says market pricing implies the incident may be viewed negatively and could weigh on investor confidence and valuation ^[18]. The site did not publish a price or name a market. The costs that can be identified so far are operational and legal, and the legal one sits in a San Francisco courtroom. Nothing in public reporting ties that lawsuit to the three departures ^[11].

I'd be wrong about disclosure being the main exposure if the next incident reaches the public from OpenAI first, inside days, through its own channels. The Sept. 16 flag process exists to produce that outcome ^[9].