computer-use-agents
Warn
Audited by Socket on May 12, 2026
1 alert found:
AnomalyAnomalysub-skills/perception-reasoning-action-loop.md
LOWAnomalyLOW
sub-skills/perception-reasoning-action-loop.md
No definitive malware behavior (e.g., credential theft, persistence, hidden payloads, explicit exfiltration destinations) is visible in the provided fragment alone. However, it is a high-impact dual-use component: it captures full-screen contents (privacy/secret exposure risk) and executes arbitrary mouse/keyboard actions from an unvalidated action dict (risk of unintended or malicious UI operations). Because the snippet is truncated and omits the model/API request/response handling, the likelihood of screenshot transmission and any safety controls cannot be fully verified from this fragment.
Confidence: 100%Severity: 60%
Audit Metadata