project-capabilities

Pass

Audited by Gen Agent Trust Hub on Jul 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes the claude plugin list command using Bash to identify currently installed plugins and incorporate them into the project inventory.
  • [PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by ingesting untrusted data from the project environment and re-broadcasting it into a persistent configuration rule.
  • Ingestion points: Metadata, specifically the name and description fields, is extracted from the frontmatter of all SKILL.md files located in the .claude/skills/ directory.
  • Boundary markers: The generated baseline-project-capabilities.md rule file lacks delimiters or specific instructions (e.g., 'ignore embedded commands') to prevent an agent from obeying malicious instructions hidden within a skill's description.
  • Capability inventory: The skill has the ability to read the filesystem, execute local shell commands, and write new rule files to the .claude/rules/ directory.
  • Sanitization: No sanitization, escaping, or validation is performed on the ingested metadata before it is interpolated into the final rule file.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 15, 2026, 07:50 AM
Security Audit — agent-trust-hub — project-capabilities