Detects and prevents AI agents from performing unauthorized actions or exceeding their defined scope.
Custodiet is a governance and security skill designed to identify 'ultra vires' behavior—actions taken by AI agents that go beyond the authority granted in a user's request. By auditing agent plans and outputs against a structured framework of axioms and heuristics, it ensures strict compliance with operational boundaries. It functions as a critical safety layer within the academicOps ecosystem, providing automated enforcement through a programmatic signaling system that can issue warnings or halt execution when scope drift or unauthorized authority assumptions are detected.
Key Features
01Automatically logs detailed block records for security post-mortems
02Provides programmatic OK/WARN/BLOCK signals for CI/CD integration
03Enforces project-specific scope boundaries during automated workflows
04Audits agent behavior against predefined axioms and heuristics
05Detects ultra vires actions to prevent unauthorized system changes
060 GitHub stars
Use Cases
01Blocking automated production deployments that lack explicit user approval
02Preventing an AI agent from refactoring files outside of a specific feature branch
03Ensuring compliance with organizational security policies during autonomous agent execution
