agent-shared-values-patterns
Pass
Audited by Gen Agent Trust Hub on Oct 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents a pattern for building agent system prompts by concatenating content from multiple external Markdown files (e.g., CLAUDE.md hierarchy).
- Ingestion points: The Python load_rule_set function and the TypeScript buildParallelPrompts function load instructions from local file paths and shared context stores into the agent's system prompt.
- Boundary markers: The build_system_prompt function uses standard Markdown headers (e.g., ## Global Rules) to delimit sections, which offers minimal protection against adversarial instructions embedded in the source files.
- Capability inventory: The provided code examples demonstrate string manipulation and local file reading; while these are not inherently dangerous, the resulting prompts are intended to govern agents that may possess privileged capabilities.
- Sanitization: The skill lacks logic to sanitize, escape, or validate the content loaded from rule files before it is interpolated into the final prompt block.
Audit Metadata