integrate-atlas-chat

Pass

Audited by Gen Agent Trust Hub on Sep 18, 2026

Risk Level: SAFEREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill fetches Python code from the official Cognite API (api.cognitedata.com) and executes it within a Pyodide runtime in the user's browser. This is the core mechanism for running server-defined tools defined within the vendor's platform.
  • [EXTERNAL_DOWNLOADS]: Downloads the Pyodide runtime from the well-known jsDelivr CDN and installs the cognite-sdk and other dependencies from standard registries using micropip.
  • [DYNAMIC_EXECUTION]: Uses the Pyodide runtime to dynamically execute Python code by constructing wrappers around tool logic and arguments at runtime using string templates.
  • [CREDENTIALS_UNSAFE]: Extracts the active Cognite session token from the client headers and injects it into the Python environment variables (COGNITE_TOKEN) to authorize SDK calls within the tool execution context. While sensitive, this facilitates the intended vendor functionality.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied messages and application state (appContext) which are passed to the agent API. The skill uses runtime argument validation for client-side tools via TypeBox to mitigate malformed tool inputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 18, 2026, 02:14 PM
Security Audit — agent-trust-hub — integrate-atlas-chat