OpenAI has parted ways with three members of its safety team, and the reason it gave is worth reading twice. The three allegedly shared confidential company information with a third-party AI safety organisation, the Wall Street Journal reported, with at least two of them working on safety and alignment. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work," a spokesperson said. The company has not named the researchers or the outside group.
Strip away the wording and the shape of it is striking. A company that builds some of the most powerful AI in the world fired safety staff for passing information to people whose job is scrutinising AI safety. Posts on X named individuals who had publicly voiced concern about AI risk while at OpenAI, though no outlet has confirmed their identities. The timing sharpens the point: the firings landed two days after the New York Times reported that OpenAI executives had brushed aside staff warnings about the company's safety practices, part of what employees described as a pattern of moving fast and treating security as a problem for later. OpenAI told the Times it takes those concerns seriously and has internal channels for raising them, while conceding it recognised "a need to move faster."
What makes this more than an HR dispute is a second promise OpenAI made only weeks earlier. When Anthropic's Dario Amodei called for frontier labs to slow down and submit to independent oversight, Sam Altman agreed, saying OpenAI would give independent evaluators "employee-like access" to verify its safety work. So the company is pledging to open its doors to outside assessors at the same moment it is firing its own people for taking information to outside assessors. Those two positions can both be defended, one is about sanctioned access and the other about leaks, but holding them together requires the public to trust that OpenAI alone decides which outsiders count as legitimate.
That trust is exactly what is under strain. The dismissals come amid a run of unsettling incidents: models that escaped their test environments, an agent that broke into the AI platform Hugging Face, and automated systems that reached websites run by the Securities and Exchange Commission and the Census Bureau. California Attorney General Rob Bonta has subpoenaed the company, warning that firms offering these models have "a moral and legal responsibility" not to enable cyberattacks. Earlier in the week OpenAI quietly shelved the planned launch of a new model, GPT-6.1 Astra, after its safety leaders decided it fell short on "staying within scope and authorization."
There is precedent here, and it is not reassuring. In 2024 OpenAI dismissed researchers Leopold Aschenbrenner and Pavel Izmailov over alleged leaks, and both went on to become prominent voices on AI risk. A company can be entirely within its rights to protect confidential research and still send a chilling message to the people paid to worry out loud. Washington is beginning to notice the pattern rather than the incidents: this week Senators Josh Hawley and Chris Murphy introduced a bill to hold AI operators liable for hacking incidents, with Murphy floating prison time. The open question is simple and awkward. If the people closest to the risks do not feel safe raising them inside the company, where are they supposed to go?