cdp

Warn

Audited by Gen Agent Trust Hub on Sep 2, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONCOMMAND_EXECUTIONPERSISTENCEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill's core functionality relies on a Node.js HTTP server (sdk/repl.ts) that accepts raw JavaScript code via a POST /eval endpoint and executes it using eval().
  • While the server binds to 127.0.0.1, any local process can interact with it on the default port 9876 to execute arbitrary code.
  • The skill also uses dynamic import() in sdk/helpers.ts to load site-specific tools from the learnings/ directory based on manifest files.
  • [COMMAND_EXECUTION]: The harness provides an environment where the agent can execute any Node.js logic, including interacting with the host filesystem and spawning child processes.
  • Interaction recipes in interaction-skills/ document using node:fs for writing files and osascript via execFile for UI automation tasks.
  • [PERSISTENCE]: The scripts/setup script automatically modifies user shell configuration files (.zshrc, .bashrc, or .profile) to ensure the CLI tool is available in future sessions.
  • It appends a command to export a new directory into the user's PATH environment variable.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from any website the agent navigates to, creating a significant surface for indirect injection.
  • Ingestion points: Untrusted data enters the context via accessibility tree snapshots (axView in sdk/axview.ts) and async page signals (drainSignals in sdk/helpers.ts).
  • Boundary markers: The axView helper produces structured text but lacks explicit markers to prevent the agent from obeying instructions embedded in the web content.
  • Capability inventory: The skill possesses high-privilege capabilities including cookie manipulation, credential scrubbing, and arbitrary browser interaction.
  • Sanitization: Minimal sanitization is performed on page content before it is passed to the LLM.
  • [EXTERNAL_DOWNLOADS]: The skill uses curl in its bash CLI and documented fetch patterns in the interaction guides to interact with local and remote resources.
  • [DATA_EXPOSURE]: The tool contains logic in sdk/session.ts to scan sensitive OS directories for browser profile data to detect running Chromium instances and read their debugging ports.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 2, 2026, 05:09 PM
Security Audit — agent-trust-hub — cdp