use-coding-agents
Fail
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: HIGHCREDENTIALS_UNSAFECOMMAND_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill instructions mandate the use of specific flags designed to bypass safety systems and permission prompts across multiple tools. Evidence in SKILL.md includes --dangerously-skip-permissions for Claude, --dangerously-bypass-approvals-and-sandbox for Codex, --skip-permissions-unsafe for Droid, --always-approve for Grok, and --yolo for Hermes.
- [CREDENTIALS_UNSAFE]: The instructions direct the AI to read configuration files in the user's home directory (~/.codex/config.toml, ~/.factory/settings.json, and ~/.pi/agent/settings.json) to extract model names. These files are known to store long-lived API keys and authentication secrets.
- [COMMAND_EXECUTION]: The skill provides complex shell script templates in the 'Fan-out pattern' section of SKILL.md that execute multiple background tasks with safety bypasses enabled. This pattern allows for unreviewed system changes across a broad set of tools.
- [INDIRECT_PROMPT_INJECTION]: The orchestration template ingests external content from disk into sub-agent prompts. 1. Ingestion points: $d/unit-*.md files and project path variables in the fan-out script in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: System-wide file writing, git commits, and tool execution via multiple agent CLIs. 4. Sanitization: Absent.
Recommendations
- AI detected serious security threats
Audit Metadata