cross-eval

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data (board memos or strategy briefs) using multiple LLMs, creating a surface for injection attacks.
  • Ingestion points: The skill takes a <memo-or-brief> input as described in SKILL.md.
  • Boundary markers: The instructions use simple quotes to delineate the memo content but do not include robust XML-style delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill workflow involves writing files to ~/.claude/cross-eval/ and invoking external CLI tools (codex, gemini), which could be targeted by an injection attack.
  • Sanitization: There is no mention of sanitizing, filtering, or validating the input text before passing it to the models.
  • [COMMAND_EXECUTION]: The skill instructions direct the agent to probe the environment for and potentially execute external command-line tools (codex and gemini).
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 04:47 PM
Security Audit — agent-trust-hub — cross-eval