cross-eval
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external board memos and strategy briefs that are passed directly into prompts for multiple LLMs.
- Ingestion points: The user-supplied
<memo-or-brief>input in the/c-level-agents:cross-evalcommand is the primary entry point for untrusted data. - Boundary markers: The prompt template lacks explicit delimiters or structural markers to clearly separate the untrusted memo content from the agent's instructions, increasing the risk that instructions hidden within a memo could be obeyed by the models.
- Capability inventory: The skill reads local files, writes analysis results to
~/.claude/cross-eval/, and performs network operations by invoking external APIs (OpenAI/Codex and Gemini). - Sanitization: There is no evidence of input validation or content filtering to sanitize malicious instructions within the memos before they are processed.
Audit Metadata