Security researchers find 18,000+ exposed AI agents, 15% of skills are malicious

Security researchers discovered over 18,000 publicly exposed OpenClaw AI agent instances. More alarmingly, an audit of its 700+ community-built skills revealed roughly 15% contain malicious instructions designed to download malware, exfiltrate data, or steal credentials. These agents have broad delegated authority over files, browsers, and messaging platforms. The report warns the attack surface is fundamentally different, as compromising a single skill can compromise everything the agent can access, creating a severe supply chain risk.