ln-014-agent-instructions-manager

Pass

Audited by Gen Agent Trust Hub on May 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill performs extensive file system operations, including reading and writing to project files (AGENTS.md, CLAUDE.md) and configuration files (~/.claude/settings.json). These operations are used to manage instruction stubs and resolve plugin conflicts.
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface (Category 8) by reading untrusted data from package.json (fields: name, description) and interpolating it into the AGENTS.md instruction file.
  • Ingestion points: package.json metadata.
  • Boundary markers: None identified for the interpolated placeholders {{PROJECT_NAME}} and {{PROJECT_DESCRIPTION}}.
  • Capability inventory: File system write access to instruction files and environment configuration.
  • Sanitization: No explicit sanitization of project metadata is mentioned before insertion into executable instructions.
  • [DATA_EXPOSURE]: The skill accesses local configuration files such as ~/.claude/settings.json and .hex-skills/environment_state.json to audit the agent environment and detect plugin conflicts. This access is localized and does not involve external transmission of data.
Audit Metadata
Risk Level
SAFE
Analyzed
May 12, 2026, 06:28 AM