exploring-llm-evaluations
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted AI generation events which may contain malicious instructions designed to influence the outcome of LLM-judge or Hog-based evaluators.
- Ingestion points: Processes $ai_generation event data via the posthog:llma-evaluation-run and posthog:query-llm-trace tools.
- Boundary markers: The skill does not explicitly define delimiters or ignore instructions headers when passing generation content to evaluators in the provided examples.
- Capability inventory: Includes capabilities to execute SQL/HogQL queries (posthog:execute-sql), run LLM-based evaluations, and execute Hog code.
- Sanitization: No explicit sanitization or filtering of the content within $ai_generation properties is specified before it is scored by the evaluators.
- [DYNAMIC_EXECUTION]: The skill generates and executes domain-specific code and queries at runtime to perform evaluations and data analysis.
- Evidence: Uses posthog:llma-evaluation-test-hog and posthog:llma-evaluation-create to execute Hog source code, and posthog:execute-sql to run HogQL queries. These are standard platform features for the PostHog environment and represent the core functionality of the skill.
Audit Metadata