agent-shared-values-patterns

Pass

Audited by Gen Agent Trust Hub on Apr 8, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill describes an architecture for 'Indirect Prompt Injection' surfaces by design. It provides patterns for constructing agent system prompts by concatenating content from various filesystem paths (e.g., CLAUDE.md, project rules). While this is the intended functionality of the pattern, it creates a surface where untrusted data could influence agent behavior.
  • Ingestion points: The load_rule_set function in SKILL.md reads content from global_path, project_path, and role_path.
  • Boundary markers: The build_system_prompt function uses simple Markdown headers (## {title}) to separate rule layers but lacks explicit 'ignore embedded instructions' delimiters or protective wrappers for the interpolated content.
  • Capability inventory: The patterns are designed to govern agent behavior, including tool usage and output formatting, making the integrity of the input files critical to system safety.
  • Sanitization: No sanitization or validation of the rule file content is present in the provided Python or TypeScript examples.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 8, 2026, 08:24 AM
Security Audit — agent-trust-hub — agent-shared-values-patterns