Not for Private Gain questions OpenAI's oversight of the agent breach

The Not for Private Gain campaign — which in 2025 had gathered dozens of law professors, former OpenAI employees and Nobel laureates to oppose OpenAI’s conversion into a public-benefit corporation — published a September 2026 update carrying a much shorter roster: five lawyers from Legal Advocates for Safe Science and Technology and the advocacy group Encode AI, among them Tyler Whitmer and Nathan Calvin. The update used the summer’s autonomous-agent breach of Hugging Face as a test of whether the nonprofit that nominally still controls the company has exercised that control in practice, framing the incident as the first serious safety event since the October 2025 restructuring.

The update centred on the Safety and Security Committee, the board body charged with monitoring frontier risk. The authors argued that the committee has no dedicated staff or full-time head of its own to supervise the operating company, and that OpenAI’s published safety commitments had not kept pace with events: they noted that the Preparedness Framework was last revised in April 2025, before the models involved in the breach existed, and characterised a newer governance framework as weaker than the original pledges. They also questioned the independence of two recent board appointments, observing that the directors lead firms that are substantial OpenAI customers.

Drawing on OpenAI’s and METR’s own post-incident reports, the update contended that the episode was, in its words, “foreseeable and avoidable” — pointing to internal responders who it said linked suspicious activity to an evaluation before the attack escalated but judged that stopping it was not required. It presented these as questions the nonprofit’s forthcoming annual report should answer rather than as settled findings.

The document listed specific things it wanted that report to address: the scope of decisions the Safety and Security Committee treats as within its authority, how and when it was informed of the incident, what resources the nonprofit has to oversee the public-benefit corporation, how OpenAI would operationalise its “stop-and-assist” commitment, and whether a future public offering would preserve nonprofit control. As an advocacy publication the update carried no binding force; it is recorded here as the clearest public attempt so far to hold the post-restructuring governance structure to the standard OpenAI set for it.