cmux-ref

Pass

Audited by Gen Agent Trust Hub on Jun 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses the cmux CLI tool to interact with the environment, including reading screen content via 'read-screen' and sending input to terminals using the 'send' command.
  • [PROMPT_INJECTION]: The skill processes user-pasted topology identifiers (refs and IDs) and uses them as routing context for agent actions.
  • Ingestion points: User-pasted text blocks in chat or terminal (SKILL.md).
  • Boundary markers: None; identifiers are parsed directly from free-form user pastes.
  • Capability inventory: Full terminal and workspace interaction via cmux (read-screen, send, open, new-surface).
  • Sanitization: Identifiers are extracted using a restrictive regex ([A-Za-z0-9:-]+), which effectively prevents the injection of shell metacharacters into the CLI commands.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 13, 2026, 07:18 AM
Security Audit — agent-trust-hub — cmux-ref