maintain-verification-skill

Pass

Audited by Gen Agent Trust Hub on Sep 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to manage and execute "harness scripts" that reside within project-local verification directories. These scripts are invoked to drive application instances (servers, UIs, or CLIs) to verify feature functionality.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it processes untrusted data from the target project's source code and feature maps.
  • Ingestion points: The skill reads the project's source code and feature description files via concurrent subagents to identify feature logic and potential documentation drift (Step 2).
  • Boundary markers: The instructions explicitly restrict the subagents used for source analysis, stating they must remain read-only and are strictly prohibited from driving the application or editing files.
  • Capability inventory: The coordinator agent is capable of executing local harness scripts, driving application sessions, and writing to the verification skill's specific directory (Step 4 and Step 5).
  • Sanitization: The skill does not define specific sanitization or escaping for the ingested code; instead, it relies on the agent's reconciliation process to filter and merge recipes derived from the source code analysis.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 7, 2026, 11:13 PM
Security Audit — agent-trust-hub — maintain-verification-skill