exploring-llm-evaluations
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from external AI generations and user messages to evaluate performance.
- Ingestion points: Untrusted content enters the context via $ai_generation events and sentiment analysis strings retrieved through posthog:execute-sql and posthog:query-llm-trace.
- Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions when the agent reads reasoning strings or generation content.
- Capability inventory: The agent can execute HogQL via posthog:execute-sql and set up scheduled reports via posthog:llma-evaluation-report-create, which can send data to external targets like email or Slack.
- Sanitization: No explicit sanitization or filtering of external content is mentioned before it is processed by the agent or included in reports.
- [DYNAMIC_EXECUTION]: The skill facilitates the creation and testing of Hog code (deterministic scripts) for rule-based evaluations.
- Script generation: The agent generates Hog source code strings for tasks like JSON schema validation and token limits.
- Execution method: These scripts are executed via the posthog:llma-evaluation-test-hog and posthog:llma-evaluation-create tools.
- Context: This functionality is a core, authenticated feature of the PostHog platform intended for objective rule-based checks.
Audit Metadata