AI Agent Risks Explode: 21,000 Deployments Exposed

The open-source AI agent, OpenClaw, has experienced a dramatic surge in public deployments, rapidly increasing from approximately 1,000 to over 21,000 instances within a single week as tracked by Censys. Security analyses reveal significant vulnerabilities, including a critical remote code execution flaw (CVE-2026-25253) and a command injection vulnerability (CVE-2026-25157). A review of the ClawHub marketplace found 7.1% of skills contained critical security flaws, exposing sensitive credentials in plaintext.

Further investigations by Wiz uncovered that Moltbook, an AI social network built on OpenClaw, left its Supabase database publicly accessible. This breach exposed 1.5 million API authentication tokens, 35,000 email addresses, and private messages containing plaintext OpenAI API keys. Concerns are growing as Meta has been observed testing OpenClaw integration, and startups are promoting wrappers for the viral project.

To address these escalating risks, Cloudflare has released the Moltworker framework. This open-source solution utilizes ephemeral containers within a Cloudflare Sandbox to isolate AI agents, preventing direct access to corporate networks and files. It employs encrypted R2 storage for persistence and Zero Trust authentication for the admin interface, offering a secure evaluation path distinct from risky local deployments on employee machines.