nightly-eval-investigation

Pass

Audited by Gen Agent Trust Hub on Nov 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from evaluation results stored in a Google Cloud Storage bucket (gs://guidance-evals/). Specifically, the scripts/investigate.ts script extracts error messages and check IDs from evals.json files and includes them in a generated markdown report. Because the agent is subsequently instructed to perform qualitative deep-dives on this report, malicious content in the evaluation output could attempt to influence the agent's diagnostic reasoning.
  • Ingestion points: Reads evals.json from GCS in scripts/investigate.ts.
  • Boundary markers: The generated report uses standard markdown headings but lacks specific delimiters to isolate untrusted error strings from the agent's instructions.
  • Capability inventory: The skill can read/write local files, execute shell commands via execSync (gcloud, node, gh), and create GitHub issues.
  • Sanitization: The script performs minimal sanitization, extracting raw message strings for inclusion in the report.
  • [COMMAND_EXECUTION]: The skill relies on child_process.execSync to perform its core functions. It executes gcloud storage commands to list and synchronize evaluation results from GCS, and gh CLI commands to create parent and sub-issues on GitHub, manage labels, and integrate with GitHub Projects.
  • Evidence: scripts/investigate.ts runs gcloud storage ls and gcloud storage rsync.
  • Evidence: scripts/publish_report.ts runs gh issue create, gh label create, gh project item-add, and gh api graphql.
  • Context: These commands are necessary for the skill's workflow and target the author's official infrastructure (GoogleChrome GitHub organization and associated GCS buckets).
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 15, 2026, 04:23 AM