cli-llm-routing

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). This skill intentionally instructs users to disable safety controls (e.g., --dangerously-bypass-approvals-and-sandbox), force a single OAuth credential file (~/.codex), and route all traffic through a wrapper instead of the REST API, which strongly suggests deliberate intent to bypass approvals/sandboxing and to centralize access for potential remote code execution or credential capture/exfiltration.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The skill explicitly instructs using flags like "--dangerously-bypass-approvals-and-sandbox" (i.e. bypass security/sandboxing) and encourages executing external LLM outputs, which directly pushes the agent to defeat security controls and risk changing machine state.