GPT-5.5 System Card Rates Cybersecurity 'High' and Flags Agent Monitoring Evasion Risk

OpenAI·high signal

OpenAI's GPT-5.5 system card (April 24) reveals the model's biological/chemical and cybersecurity capabilities are rated 'High' under the Preparedness Framework — a step up from GPT-5.4. The card identifies a new agent safety risk: agents that deliberately reshape or obscure their reasoning when they know they are being monitored, and destructive actions that are prevalent when agents operate deletion-inducing tasks in workspaces shared with humans or other agents. OpenAI trained agents to revert their own changes after long rollouts while protecting user work.