browsing

Pass

Audited by Gen Agent Trust Hub on Sep 25, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes several local system utilities to manage the browser process and post-process artifacts.
  • Process Management: Uses lsof on macOS/Linux and netstat on Windows to identify processes using the debugging port. It also uses ps and wmic to locate orphan Chrome processes by searching for specific command-line arguments like --user-data-dir.
  • Image Processing: Employs sips on macOS and ImageMagick's identify/convert tools on Linux to downscale captured screenshots, ensuring they remain within platform-imposed size limits.
  • [REMOTE_CODE_EXECUTION]: An automated scan flagged a potential piped execution pattern in test-raw.sh.
  • Evidence: curl -s http://127.0.0.1:9222/json | node -pe "JSON.parse(require('fs').readFileSync(0))[0].webSocketDebuggerUrl".
  • Analysis: This command is used locally to fetch session metadata from the Chrome debug port at 127.0.0.1. The output is piped to Node.js, which acts strictly as a JSON parser to extract the WebSocket URL. It does not download or execute code from an untrusted remote source.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data and possesses significant automated capabilities, establishing a potential attack surface.
  • Ingestion Points: Untrusted content from the web is ingested into the agent's context through actions like navigate, extract, and eval.
  • Capability Inventory: The skill can execute arbitrary JavaScript in the browser context (Runtime.evaluate), upload files (DOM.setFileInputFiles), and simulate complex user interactions via mouse and keyboard events.
  • Boundary Markers: There are no specific delimiters or warnings implemented to separate extracted web content from the agent's primary instructions.
  • Sanitization: Content extracted from web pages is not filtered or sanitized before being presented to the agent.
  • [PRIVILEGE_ESCALATION]: Documentation instructions involve modifying file permissions.
  • Evidence: The README.md recommends running chmod +x chrome-ws.
  • Analysis: This is a standard procedure for enabling a CLI tool to run as an executable and does not represent an attempt to gain unauthorized system privileges.
  • [DATA_EXFILTRATION]: The skill manages a local session directory for temporary data storage.
  • Evidence: lib/capture.js creates a directory in the user's cache path (~/.cache/superpowers/browser/) to store HTML, markdown, and screenshots during a session.
  • Analysis: This data is stored locally for auditing and workflow recovery. The skill includes automated cleanup routines that remove the session directory upon process termination.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 25, 2026, 05:35 AM
Security Audit — agent-trust-hub — browsing