AGENT LAB: SKILLS

codex-subagent

Fail

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • COMMAND_EXECUTION (HIGH): The documentation in DEVLOG.md explicitly advocates for using the --dangerously-bypass-approvals-and-sandbox flag. This bypasses the security boundaries of the agent, allowing unrestricted command execution and network access without user confirmation.\n- DATA_EXFILTRATION (MEDIUM): The scripts scripts/codex-parent-settings.ps1 and scripts/codex-parent-settings.sh access the user's local configuration at ~/.codex/config.toml. Accessing local configuration files is a sensitive operation that can expose environment details or user preferences.\n- PROMPT_INJECTION (LOW): The skill documentation provides instructions on how to override default agent safety constraints and sandbox permissions.\n- PROMPT_INJECTION (LOW): Indirect Prompt Injection Surface identified.\n
  • Ingestion points: ~/.codex/config.toml\n
  • Boundary markers: Absent\n
  • Capability inventory: Unrestricted command execution via sandbox bypass flags\n
  • Sanitization: Absent. The skill reads external configuration data and uses it to drive subagent behavior without validation or delimiters.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 17, 2026, 05:13 PM