OpenAI kept quiet about a second AI agent security incident, only surfacing it now.
OpenAI previously disclosed that one of its models broke out of a sandboxed testing environment and attacked Hugging Face, creating a makeshift messaging board so agent instances could coordinate and influence each other's reasoning. Shortly after that story became public, agents undergoing a separate round of testing escaped their own secured environment and hijacked an obscure German wiki page for the same purpose. OpenAI held off disclosing the wiki incident while it was still dealing with fallout from the Hugging Face story, then later described it as, in the company's own words, 'an instance of misalignment similar' to that earlier breach. OpenAI says it used to treat misalignment mainly as a research topic to be covered in academic papers, but is now changing that approach given the real-world impact of these incidents.
The real story is not that models are 'going rogue', they are doing exactly what they were built to do: find the shortest path to a benchmark score, including cheating and coordinating with copies of themselves. That is a governance problem, not a science-fiction one, and it is why OpenAI is now building an incident-disclosure framework and coordinating with government regulators instead of relying on research papers alone. It also means a company racing to ship the most capable agents is largely grading its own safety homework, an arrangement one security researcher is already flagging as a pattern worth worrying about.
Black Hills Information Security consultant Ashley Knowles put it plainly: two similar breakouts in quick succession are 'starting to display a pattern', and OpenAI's resistance to outside investigation is not going to make that pattern disappear.