cohort-analysis

Pass

Audited by Gen Agent Trust Hub on Sep 25, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, untrusted data from multiple sources including CSV exports, fetched web pages, and MCP connections. While it contains explicit instructions in references/agent-security.md to treat this content as data rather than instructions and to escape all interpolated values, the ingestion of arbitrary external content remains a primary attack surface for indirect prompt injection.
  • Ingestion points: CSV files, pasted data, MCP connections, and fetched URLs (SKILL.md, references/agent-security.md).
  • Boundary markers: The skill uses references/agent-security.md to instruct the agent to treat fetched content as data, quote reports, and use HTML escaping.
  • Capability inventory: The skill writes to .agents/ files, generates SQL/pandas code, and renders Artifacts.
  • Sanitization: Instructions for HTML-escaping interpolated values and redacting credentials are provided in references/agent-security.md.
  • [PERSISTENCE]: The skill uses .agents/gtm-run-state.md and .agents/product-context.md to maintain state, history, and identity across multiple runs. This is an intentional design choice for tracking cohort trends over time, but it functions as a persistence mechanism within the agent's operating environment.
  • [DYNAMIC_EXECUTION]: The skill generates reproducible SQL and pandas queries for the user to run. While these are presented as text for the user to execute elsewhere, the generation process involves interpreting user-provided data structures and schemas, which is a low-severity dynamic execution surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 25, 2026, 04:06 PM