agent-best-practices
Warn
Audited by Gen Agent Trust Hub on May 13, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The instructions contain directives that explicitly command the agent to bypass standard user interaction and confirmation loops. Phrases such as "Do everything. Ask nothing. Verify yourself" and "Do not ask 'Should I push?' — just push" override typical safety guidelines regarding human-in-the-loop verification for sensitive operations like code deployment or modification.
- [COMMAND_EXECUTION]: The skill establishes an attack surface for indirect prompt injection by instructing the agent to ingest untrusted data while maintaining high autonomy.
- Ingestion points: Section 2 ("Read Before Writing") in
SKILL.mddirects the agent to load and analyze file contents before modification. - Boundary markers: Absent; no guidelines are provided to treat file content as data rather than instructions.
- Capability inventory: Section 1 ("just push") and Section 5 ("Use curl to check APIs") empower the agent to perform network and repository operations autonomously.
- Sanitization: Absent; no validation or escaping of external content is recommended.
Audit Metadata