social-sandbox

Warn

Audited by Gen Agent Trust Hub on Jun 27, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The Python recipe provided in the skill dynamically modifies the system path to load local modules from a directory in the user's home folder (~/.claude/skills/persona-sim). This pattern allows the execution of code from outside the immediate skill directory and relies on the security of the target directory.
  • [PROMPT_INJECTION]: The skill processes a "scenario" variable by directly interpolating it into a task string for LLM-simulated personas. This creates an indirect prompt injection surface where instructions embedded in the scenario could influence or override the intended simulation behavior.
  • Ingestion points: The scenario variable defined in the Python script in SKILL.md.
  • Boundary markers: None identified; the scenario is directly concatenated into the prompt string.
  • Capability inventory: The script uses a custom generate function and has the ability to load external library files via path manipulation.
  • Sanitization: No validation or escaping is applied to the scenario text before it is used in the prompt.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 27, 2026, 07:09 AM
Security Audit — agent-trust-hub — social-sandbox