maintain-agentic-system
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the target repository during the 'Evidence Scan' and 'Region Merge' gates, creating a surface for indirect prompt injection.
- Ingestion points: The skill reads repository-local agent files ('MINE'), changelogs, and manifest files to detect changes and provenance (referenced in
contracts/delta-audit.mdandcontracts/modes-and-scope.md). - Boundary markers: The skill includes 'Non-Negotiable Principles' in
SKILL.mdthat explicitly limit scope to agent-system files and strictly forbid reading session data. It also mandates producing a maintenance plan for user approval before any writes. - Capability inventory: The skill has the capability to write and modify files in the repository during Gate 5 (Maintenance Plan And Approved Apply).
- Sanitization: A formal three-way merge logic (detailed in
contracts/merge-model.md) is employed to highlight differences between the baseline, templates, and repository files, requiring manual resolution of conflicts by the user.
Audit Metadata