antithesis-triage

Pass

Audited by Gen Agent Trust Hub on May 14, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill makes extensive use of the agent-browser CLI to interact with the Antithesis web interface. It executes shell commands to manage browser sessions, inject a custom JavaScript runtime (antithesis-triage.js) into the browser context, and download log files using the download-logs.sh script. It also uses jq for processing structured JSON data.
  • [EXTERNAL_DOWNLOADS]: The skill requires several prerequisites including snouty, agent-browser, and jq. The SKILL.md file provides links to the official GitHub repositories (e.g., github.com/antithesishq/snouty, github.com/vercel-labs/agent-browser) and official project pages for these tools. These sources are considered trusted or well-known services.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize external data from Antithesis reports and application logs generated by the system under test (SUT).
  • Ingestion points: Data is extracted from the Antithesis web UI via a JavaScript runtime and from downloaded JSON log files.
  • Boundary markers: The instructions do not use explicit delimiters (like XML tags or markdown blocks) with 'ignore embedded instructions' warnings when processing raw log content.
  • Capability inventory: The skill has the capability to execute shell commands (agent-browser, jq, cat), perform file system operations (via agent-browser download), and navigate to network URLs.
  • Sanitization: A Python script (process-logs.py) is provided to strip ANSI escape codes from log output before analysis.
Audit Metadata
Risk Level
SAFE
Analyzed
May 14, 2026, 01:45 PM