architecture-hardening-loop
Pass
Audited by Gen Agent Trust Hub on Oct 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including user-supplied source code, commit messages, and scanner reports, while maintaining significant capabilities like file modification and tool execution.\n
- Ingestion points: Source code files in the user-defined scope, commit history/messages via Git, and diagnostic output from the
improve-codebase-architecturetool.\n - Boundary markers: Uses a 'Hardening Contract' to lock scope and invariants, and provides explicit instructions to treat user-provided attachments as data only, ignoring any embedded commands.\n
- Capability inventory: The agent can perform file writes to apply architectural fixes, execute shell commands for testing/verification, and invoke other integrated skills.\n
- Sanitization: Includes logic to specifically exclude secrets, credentials, and local-only paths from the scanning and modification scope.\n- [EXTERNAL_DOWNLOADS]: The skill relies on external components not authored by the primary vendor, specifically
improve-codebase-architectureandcodebase-designfrom themattpocock/skillsrepository. These are essential for the scanning and design enforcement phases of the loop.
Audit Metadata