codex-ab

Pass

Audited by Gen Agent Trust Hub on Jul 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Executes a local bash script (scripts/codex-panel-ab.sh) which calls git for repository metadata and codex for AI-assisted code review passes.
  • [EXTERNAL_DOWNLOADS]: Requires the codex CLI tool to be installed on the system; its presence is verified during the preflight stage.
  • [DATA_EXFILTRATION]: Accesses project code and Git history to generate diffs for analysis. Results are written locally to the project's .claude/reviews/ directory.
  • [PROMPT_INJECTION]: Demonstrates an indirect prompt injection surface because it passes untrusted code from Git diffs to an external AI model (codex).
  • Ingestion points: Git diff content processed in scripts/codex-panel-ab.sh via the codex exec command.
  • Boundary markers: Includes 'GUARD' instructions in the prompt to restrict the AI's behavior and output format.
  • Capability inventory: Shell command execution and local file system writes.
  • Sanitization: Employs a read-only sandbox flag (-s read-only) for dimension-focused passes to mitigate risks from processing potentially malicious code comments.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 21, 2026, 03:50 AM
Security Audit — agent-trust-hub — codex-ab