agentic-loops
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill architecture relies on a persistent state file to persist the agent's goal and planned actions across iterations. This creates a surface where the agent's future behavior is guided by data written in previous cycles.\n
- Ingestion points: The agent reads the state from files such as templates/loop-state.json and examples/debug-loop.json, and updates them as instructed in SKILL.md.\n
- Boundary markers: The references/loop-state-contract.md defines strict transition rules, invariants, and mandatory budget checks.\n
- Capability inventory: The skill utilizes python3 for state validation (scripts/validate_loop_state.py) and permits tool calls within a bounded scope.\n
- Sanitization: The Python script scripts/validate_loop_state.py performs rigorous schema validation, type checking, and enforces monotonicity of spend metrics.\n- [DYNAMIC_EXECUTION]: The test suite scripts/test_validate_loop_state.py uses importlib.util to dynamically load and execute the validator logic from scripts/validate_loop_state.py. This dynamic loading is used for test execution and is restricted to the local files bundled with the skill.
Audit Metadata