OpenClaw releases a security dataset for agent skills

OpenClaw released ClawHub Security Signals, a dataset and accompanying paper built from multiple security scanners applied to a snapshot of the public ClawHub agent-skill catalog. The project is designed to help researchers study how VirusTotal, static analysis, and semantic-risk tools disagree when evaluating skills that can extend an agent's capabilities. The authors stress that its labels are silver-standard automated signals rather than verified judgments: a suspicious result means a skill deserves review, not that it is malicious. The release also redacts secrets and excludes private artifacts, while warning that automated sanitization is imperfect. This is a useful security artifact because agent skills combine instructions, code, permissions, and external services in ways that traditional package scanning does not fully capture.

Key details: June 1, 2026 snapshot, Multi-scanner dataset, Agent-skill security, Silver-standard labels, VirusTotal and static-analysis signals, Open dataset.

Continue swiping for more AI Brief stories.