anchor-browser-automation

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (HIGH): The skill creates a high-risk surface by processing untrusted data (web page content) and providing tools with side effects.
  • Ingestion points: External data from the browser is ingested via the Anchor Browser toolkit.
  • Capability inventory: Includes RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH, allowing for complex actions and potential code execution.
  • Boundary markers: None identified in the skill instructions to prevent the agent from obeying instructions found on web pages.
  • Sanitization: No evidence of sanitization or filtering of external content before it reaches the LLM context.
  • Remote Code Execution (MEDIUM): The skill utilizes RUBE_REMOTE_WORKBENCH and RUBE_MULTI_EXECUTE_TOOL to execute operations on a remote server. The lack of clear constraints on the run_composio_tool() function within the workbench could lead to arbitrary code execution if compromised.
  • External Dependency (LOW): The skill requires connecting to an external MCP endpoint (https://rube.app/mcp). While typical for MCP, this source is not on the pre-approved trusted list, necessitating manual verification of the provider's security posture.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 06:51 AM