autonomous-fixes
Pass
Audited by Gen Agent Trust Hub on Mar 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill leverages the
Bash,Write,Edit, andAgenttools to perform code remediation and verification. This functionality is supported by a setup script (scripts/setup-hook.sh) that modifies the global~/.claude/settings.jsonconfiguration file to install security hooks and configure the model toclaude-opus-4-6. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the processing of untrusted finding data.
- Ingestion points: Findings are parsed from markdown files in the
docs/_autonomous/directory (Phase 1). - Boundary markers: The
references/finding-parser.mddoes not specify the use of delimiters or 'ignore' instructions for the content being parsed. - Capability inventory: The skill has full capability to modify source code and execute shell commands through the Orchestrator-Agent delegation model (Phase 3).
- Sanitization: Findings are processed and used to generate fix plans without documented validation or sanitization of the input text.
Audit Metadata