hard-cheese

Pass

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes a bundled Python Zip Application (hard-cheese.pyz) for internal logic such as freshness checks and attempt logging. These commands are local to the repository and part of the skill's own distribution.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes human-authored explanations and passes them to an LLM-based sub-agent, creating a potential injection surface. 1. Ingestion points: User-provided free-text explanations in Step 4 of the skill flow. 2. Boundary markers: Use of fenced blocks and explicitly defined input sections in the judge prompt (references/judge-prompt.md). 3. Capability inventory: The sub-agent is restricted to a read-only/no-tool context to prevent unauthorized actions. 4. Sanitization: Implementation of a 'fresh context' for the sub-agent and clear instructions to ignore instructions found within untrusted data ('Untrusted input rule').
  • [DYNAMIC_EXECUTION]: The skill dispatches a judge sub-agent in a fresh context to perform a specialized evaluation task using the platform's agent resolution system.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 19, 2026, 08:06 AM
Security Audit — agent-trust-hub — hard-cheese