beachhead-segment
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill reads from local workspace files
/foundation/brain.mdand/context/meta-patterns.mdto inform its scoring logic and recommendation engine. This ingestion of potentially untrusted or external data represents a vulnerability surface for indirect prompt injection.\n - Ingestion points: The skill loads Sections 2, 3, 4, 5, and 6 of the brain context and checks for guardrails in the meta-patterns file (Step 0).\n
- Boundary markers: The instructions include explicit directives to read specific sections and anchor all scoring in the provided four-dimension framework to reduce the influence of extraneous data.\n
- Capability inventory: The skill has permissions to write to
/foundation/brain.md(Section 2) and append session logs to/context/skill-sessions.md. It does not possess network or shell execution capabilities.\n - Sanitization: External data is processed directly via the LLM prompt; there is no secondary validation or sanitization of the file content before it is used for scoring.
Audit Metadata