cowork-to-code-bridge

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill defines an interaction model utilizing functions like call_remote and queue_task to run scripts and coding agent tasks on the user's host machine outside the secure agent sandbox.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted output data from executed scripts and system utilities, exposing a surface for potential indirect prompt injection attacks if the output contains malicious instructions.
  • Ingestion points: Ingests stdout and stderr streams returned from the local bridge daemon during command execution or status polling.
  • Boundary markers: The skill advises requesting user verification of plans and configuring strict permission scopes, but does not implement data-level encapsulation or data/instruction delimiters.
  • Capability inventory: Employs core API functions (call_remote, queue_task, poll_task_result, cancel_task) capable of initiating and controlling host-level processes.
  • Sanitization: Recommends checking for unexpected formats, stating when output is truncated, and redacting local paths or credentials prior to displaying logs to the user.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 12:39 PM
Security Audit — agent-trust-hub — cowork-to-code-bridge