staging-fix
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from staging logs, error reports, and user-provided reproduction steps (SKILL.md Workflow Step 1). This creates a potential surface for indirect prompt injection if malicious instructions are embedded in the logs. Evidence:
- Ingestion points: Error reports, pasted logs, and reproduction steps (SKILL.md Step 1).
- Capability inventory: Local test execution, git push operations, and GitHub PR management (SKILL.md Steps 2-3).
- Sanitization: Mandatory redaction of tokens, keys, cookies, and identifiers before output (SKILL.md Rules).
- Boundary markers: The skill emits phase transitions (Stage / Found / Next / Needs user), though it lacks specific data delimiters for log content.
- [COMMAND_EXECUTION]: The workflow utilizes standard command-line tools including git and the GitHub CLI (gh) to push changes and manage Pull Requests (SKILL.md Step 3). All mutating remote actions are strictly protected by a mandatory combined user approval requirement before execution.
- [DYNAMIC_EXECUTION]: The skill generates and executes local tests to verify bug fixes before deployment (SKILL.md Step 2). This activity is restricted to the local development environment and is a standard component of the bug-fixing workflow.
Audit Metadata