Cybersecurity — 2026-04-22

SecurityScorecard finds 28,663 internet-exposed OpenClaw control panels, three high-severity CVEs with public exploits

SecurityScorecard published a scan identifying 40,214 internet-exposed instances of the OpenClaw agentic-AI framework, with 28,663 unique IPs hosting control panels reachable from the public internet, TechRadar reported April 21. Roughly 63 percent of observed deployments are vulnerable to remote code execution requiring no user interaction, and three high-severity CVEs (CVSS 7.8 to 8.8) have public exploit code already circulating. The Peter Steinberger framework reached 346,000 GitHub stars faster than any project in GitHub history; researchers note that an AI agent with full machine access exposes that same access to anyone who can compromise it. CrowdStrike, Reco, and Barracuda have published parallel advisories.

Analysis
The failure mode here is ungoverned-software-rollout, not library vulnerability, the same default-exposure pattern as early Kubernetes dashboards, corroborated across SecurityScorecard, CrowdStrike, Reco, and Dark Reading. The scale of publicly-reachable control panels almost certainly translates to mass automated exploitation within 30 days as commodity scanners pivot to OpenClaw signatures; the compounding risk is that an exploited agent carries lawful enterprise SaaS credentials, pushing the blast-radius analogy past Mirai's dumb-device botnet toward something with authenticated access to production systems. Agentic-AI frameworks now constitute an attack-surface category with no clean CISA playbook mapping, and the security ecosystem's response to the first operational stress test of that category is lagging badly.

4 sources

OpenClaw Trojan Horse AI agents give hackers full control of 28,000+ systems - TechRadar
What Security Teams Need to Know About OpenClaw, the AI Super Agent - CrowdStrike
OpenClaw: The AI Agent Security Crisis Unfolding Right Now - Reco
Critical OpenClaw Vulnerability Exposes AI Agent Risks - Dark Reading