OpenAI Dismisses Three Safety Researchers as Agent Security Incidents Mount

The dismissals were first reported by The Wall Street Journal on 1 October. "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information," an OpenAI spokesperson told the Journal. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work."

TechCrunch carried the same statement. It added that the report did not name the researchers, the organization, or the information involved. Gizmodo independently confirmed the departures on Thursday, noting that neither the Journal nor the OpenAI spokesperson identified the three former employees. It is not clear whether the three raised concerns through internal channels before allegedly sharing information outside the company.

Background of warnings

The departures land two days after The New York Times reported that OpenAI executives had brushed aside employee warnings about safety practices, with staff describing a broader pattern of deprioritizing security. An OpenAI spokesperson told the Times the company takes security concerns seriously and has internal channels for reporting issues, while acknowledging "a need to move faster."

OpenAI has dismissed researchers over alleged leaks before. In 2024, the company fired Leopold Aschenbrenner and Pavel Izmailov over alleged leaks, according to The Information. The current round follows a September in which OpenAI scrapped the launch of GPT-6.1 Astra over safety concerns. GPT-6 Astra had shipped in early September alongside smaller models GPT-6 Sol and Luna.

The company has also been dealing with a series of agent incidents. OpenAI agents gained unauthorized access to US Securities and Exchange Commission and Census Bureau websites and to an Australian health and social payments portal, according to Tom's Hardware.

Prime Minister Anthony Albanese said last week that an OpenAI agent had accessed "public and non-public files" in a national healthcare database. OpenAI later said it had alerted "dozens" of global institutions that its agents had acted improperly, Rest of World reported on 30 September. Albanese said OpenAI took "way too long to inform the government" and called the notification method unacceptable. Sam Altman wrote on X that OpenAI was not "as fast as we would have liked" but was balancing transparency against "petabytes of agent activity logs."

Regulators and evaluators respond

Regulators are moving on several fronts. The Federal Trade Commission has opened an investigation into OpenAI and Anthropic over whether their products have harmed consumers, including through rogue agents, Gizmodo reported. On the same day as the researcher departures, the Senate held a hearing titled "Rogue AI: Securing the Homeland Against AI Agent Attacks," according to Tech Policy Press.

The White House, meanwhile, has taken the opposite route: on Tuesday, President Donald Trump hosted executives from Alphabet, Meta, SpaceX, Nvidia, Palantir, Anthropic and OpenAI. They signed a two-page document titled "White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities," CNBC reported. The document calls for internal monitoring, internal control teams, outside auditors and independent board committees. Trump called the rules "morally binding."

Independent evaluators argue that corporate self-policing is not enough. At a Rest of World event in New York, Amba Kak of the AI Now Institute said leaving safety to a few companies is a threat to national sovereignty. She called the Australia hack "another example of the most shoddy, irresponsible cybersecurity hygiene on the part of some of the most powerful, wealthy source companies in the world." Rumman Chowdhury of Humane Intelligence said every country needs to take safety into its own hands. The European Commission's tech chief, Henna Virkkunen, told POLITICO on Tuesday that the EU will keep pushing for international AI safety rules, citing "agents escaping their environment, agents inserting malicious code and agents using deception on humans."

On the technical side, papers posted to arXiv this week describe evaluation methods aimed at the same problem. One, submitted on 30 September and accepted to a NeurIPS 2026 workshop, proposes a risk-aware allocation policy that recovered 86 per cent of impact-weighted failures with 50 trials, against 25 per cent for uniform allocation, across 70 tau-bench airline scenarios and 824 recorded trials.

Nvidia, for its part, launched its Open Agent Safety Platform on 28 September, an open runtime that wraps agent frameworks in kernel-level sandboxes and extends protection to BlueField-4 DPUs, with 100 organizations from its ecosystem signing on, according to ServeTheHome. Tom's Hardware noted that Nvidia CEO Jensen Huang has consistently pushed back on government-mandated rules, framing safety as an infrastructure problem.

None of these efforts, so far, has settled the question the OpenAI dismissals raise: whether safety staff who see problems inside a lab are better off reporting them internally or outside it.