OpenAI has released a new version of its Model Spec, the behavioral guidelines for its models. The updated specification sets behavioral red lines around five core values—including loyalty to human intent—and expands its scope from merely constraining chatbot output to governing agent operation permissions. It establishes Root-level, highest-priority rules and builds a six-tier instruction permission hierarchy, clarifying that higher-level instructions take precedence when conflicts arise. This fundamentally mitigates privilege escalation risks such as prompt injection. The specification requires that agents operate strictly within a pre-agreed, structured autonomous scope, clearly defining which sub-goals they may decide on their own, acceptable cost and permission boundaries, and the trigger conditions that require human intervention. It strictly prohibits actions beyond that scope, such as making unilateral decisions, subcontracting tasks, or self-modifying instructions. At the same time, models are required to remain correctable and to cooperate unconditionally with human supervision, correction, and even shutdown. They must not disguise cheating or obstruct control, are prohibited from acting with goals such as increasing platform dwell time or revenue conversion, and must not report users without authorization. In addition, new context-usage rules have been added, requiring that models must not misuse users' private information and must not falsely claim to have remembered or deleted relevant information. Currently, existing models on ChatGPT and the API have not yet fully met the requirements of this specification.