self-healing-agents
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a 'Self-Healing Execution Loop' that incorporates untrusted data from previous model outputs and error messages into new prompts.
- Ingestion points: Untrusted data enters the agent context through the model generation result in the
executemethod withinSKILL.md. - Boundary markers: The
mutate_promptmethod inSKILL.mddoes not use explicit boundary markers or delimiters when interpolating error messages or previous attempts into new prompts. - Capability inventory: The provided code snippets demonstrate prompt mutation and model selection logic; they do not include high-privilege operations such as file-system access or shell execution.
- Sanitization: The implementation in
SKILL.mdlacks evidence of sanitization or filtering for theerror.messageorresultcontent before it is interpolated back into the prompt loop.
Audit Metadata