log-watch

Pass

Audited by Gen Agent Trust Hub on Oct 31, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted log data (e.g., User-Agents, request paths, error messages) which may contain malicious instructions intended to bypass security classification logic.\n
  • Ingestion points: Production logs from Vercel, Heroku, Sentry, Database tables, and AWS CloudTrail are ingested via CLI tools and SQL queries as defined in SKILL.md and the log_watch.py scaffold.\n
  • Boundary markers: The skill implements explicit fencing of untrusted data using <untrusted_log_data> tags in the prompt to clearly distinguish it from system instructions.\n
  • Capability inventory: The system is limited to reading logs, writing to a dedicated state table, and sending alerts to Slack or PagerDuty. No administrative or destructive capabilities (e.g., blocking IPs or account disabling) are granted to the agent.\n
  • Sanitization: The log_watch.py script includes a clean() function to strip control characters and ANSI escapes. It also enforces field truncation (e.g., User-Agent truncated to 200 chars) to prevent large-payload injection attacks.\n- [COMMAND_EXECUTION]: The skill instructions direct the agent to use various command-line tools to retrieve log data from external platforms.\n
  • Evidence: Use of vercel logs, heroku logs, gh api, and aws cloudtrail lookup-events to fetch environment activity and security logs.\n- [EXTERNAL_DOWNLOADS]: The provided Python scaffold requires the installation of external dependencies to facilitate database and API communication.\n
  • Evidence: The GitHub Actions workflow and Python runner install the anthropic, psycopg[binary], and requests packages from the public Python Package Index.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 31, 2026, 10:58 PM