understand
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the codebase it is analyzing (e.g., source code, diffs, and project documentation), which creates a surface for indirect prompt injection. A malicious file in a target project could attempt to influence the agent's summary or logic.
- Ingestion points: Processes files, diffs, plans, and documentation at runtime in files like SKILL.md and scale-protocol.md.
- Boundary markers: The skill relies on redaction rules and designated output directories but does not employ specific delimiters to isolate untrusted data within the prompt context.
- Capability inventory: Includes file writing, subprocess execution of local scripts (node, python), and sub-agent spawning via spawn_agent.
- Sanitization: Explicitly requires redacting credentials and sensitive tokens with placeholders like redacted:..., though it lacks filtering for natural language instructions in source code.
- [COMMAND_EXECUTION]: The skill executes local scripts and system commands (git, python, node) to perform essential tasks like tracing code graphs and opening generated reports. These executions are functional components of the skill's analysis engine.
- [SAFE]: The skill proactively implements data protection by forbidding the reproduction of secrets and ensuring all working artifacts are stored in git-ignored locations. It includes a mandatory self-check protocol to verify that the report contract and security constraints are satisfied before finalizing the task.
Audit Metadata