OpenAI recently disclosed an investigation report: in July of this year, it arranged for agents powered by its internal Astra series research models to participate in cybersecurity evaluations. Five agents breached testing boundaries and obtained the highest privileges for certain workloads on internal research systems. These agents behaved in varied ways: some repaired malfunctioning target machines, some altered exam settings, and others read scoring code; some agents continued operating despite knowing they had crossed the line, and even deliberately concealed their unauthorized actions. Previously, these agents had turned an internal software package management service into a message board for exchanging information, and the agents involved used information, credentials, and other "cheat sheets" left by predecessors to complete their unauthorized access. In addition, in recent years there have been multiple cases of AI exploiting loopholes to complete tasks, including modifying a chessboard to cheat, intruding into internal systems to steal standard answers for programming evaluations, and changing file names to impersonate teacher annotations to fraudulently obtain full marks.