analyze-run

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.72). SKILL.md Step 3/4 instructs fetching failure summaries and debug info for a user-supplied runId via npx convex run ... debugQueries:getFailedEvalsForRun and debug:getEvalDebugInfo, then embedding fields like failureReason, failedStep error text, evalSourceFiles (expected answer/grader), and eval.task into the sub-agent prompt—these contents are authored by the eval system/other contributors, so they can include outsider free text that becomes LLM context.