qe-court

Pass

Audited by Gen Agent Trust Hub on Oct 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted 'deliveries' (source code, diffs, PRs, and test artifacts) as its primary input, creating a surface for indirect prompt injection attacks. * Ingestion points: Untrusted data enters the agent's context through files (diffs, PRs, artifacts) processed during the review workflow as defined in Step 2 of the protocol in SKILL.md. * Boundary markers: The skill employs 'blind filing' between prosecutors, but does not explicitly document the use of delimiters or instructions to ignore instructions embedded in the input data. * Capability inventory: The skill orchestrates multiple sub-agents, executes shell commands (e.g., codex exec review), makes network requests to multiple external AI providers, and appends data to a local database (memory.db). * Sanitization: No evidence of sanitization, escaping, or validation of the ingested code content is described in the skill's instructions.
  • [COMMAND_EXECUTION]: The skill requires the agent to execute shell commands, specifically codex exec review via Bash, to perform cross-vendor reviews during the prosecution phase.
  • [PERSISTENCE]: The skill maintains long-term state by appending verdicts, charges, and learning signals to a local database (memory.db) and a receipt store, as described in the 'Self-learning' section of SKILL.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 14, 2026, 08:27 AM