exploring-llm-evaluations

Pass

Audited by Gen Agent Trust Hub on Sep 25, 2026

Risk Level: SAFEDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill enables the creation and dry-running of "Hog" source code via posthog:llma-evaluation-create and posthog:llma-evaluation-test-hog. This code is executed within the PostHog ingestion pipeline to evaluate events. While this is the core intended purpose of the skill, it creates a surface for dynamic script execution.
  • [COMMAND_EXECUTION]: The tool posthog:execute-sql allows for the execution of arbitrary HogQL/SQL queries against the project's ClickHouse database for investigative purposes.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes AI generation data, which often contains untrusted user messages and model responses, exposing an indirect prompt injection surface.
  • Ingestion points: Generation data enters the agent context through posthog:query-llm-trace and posthog:llma-evaluation-run as described in the workflows in SKILL.md.
  • Boundary markers: Absent. The skill does not define specific delimiters or "ignore previous instruction" wrappers for the content being evaluated.
  • Capability inventory: The skill possesses the ability to execute database queries (posthog:execute-sql) and run/test custom scripts (posthog:llma-evaluation-test-hog).
  • Sanitization: Absent. There is no evidence of filtering or sanitizing event properties before they are passed into evaluation prompts or Hog code logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 25, 2026, 01:23 AM
Security Audit — agent-trust-hub — exploring-llm-evaluations