principle-encode-lessons-in-structure
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
- [NO_CODE]: The skill package consists entirely of markdown documentation and YAML configuration. No scripts, binaries, or executable code are included in the provided files.
- [INDIRECT_PROMPT_INJECTION]: The skill defines a process where the agent ingest data from human interventions and test results to create structural mechanisms (scripts, lints, runtime checks). This creates a vulnerability surface where an attacker providing malicious feedback or triggering specific test failures could influence the agent to generate and persist malicious code.
- Ingestion points: Human corrections, test failures, and unexpected outcomes encountered during agent operation (described in
SKILL.md). - Boundary markers: None. The instructions do not specify delimiters or validation steps to differentiate between legitimate learning signals and adversarial input.
- Capability inventory: The instructions explicitly encourage the agent to create scripts, lint rules, metadata flags, and runtime checks, which implies file-writing and environment modification capabilities across the workspace.
- Sanitization: No instructions are provided for sanitizing, verifying, or escaping the content derived from 'lessons' before they are encoded into structural mechanisms.
Audit Metadata