Are your developers already running OpenClaw at home? It’s time to take a closer look at the security risks associated with this open-source AI agent. Censys has tracked over 21,000 publicly exposed deployments of OpenClaw in just under a week. Bitdefender’s GravityZone telemetry has confirmed security leaders’ fears, with employees deploying OpenClaw on corporate machines with single-line install commands, granting autonomous agents shell access, file system privileges, and OAuth tokens to sensitive services like Slack, Gmail, and SharePoint.
CVE-2026-25253, a one-click remote code execution flaw rated CVSS 8.8, allows attackers to steal authentication tokens with a single malicious link, leading to full gateway compromise in milliseconds. Another vulnerability, CVE-2026-25157, enables arbitrary command execution through the macOS SSH handler. A security analysis of ClawHub marketplace skills revealed that 7.1% of the registry contains critical security flaws exposing sensitive credentials in plaintext. Additionally, roughly 17% of skills analyzed by Bitdefender exhibited malicious behavior.
The security risks extend beyond OpenClaw itself, as researchers discovered that Moltbook, an AI agent social network built on OpenClaw infrastructure, left its entire Supabase database publicly accessible, exposing 1.5 million API authentication tokens, 35,000 email addresses, and plaintext OpenAI API keys in private messages. The situation is dire, with security leaders needing a middle path between ignoring OpenClaw and deploying it on production hardware.
Cloudflare’s Moltworker framework offers a solution with ephemeral containers that isolate the agent, encrypted R2 storage for persistent state, and Zero Trust authentication on the admin interface. By following a four-step process, you can set up a secure evaluation instance for OpenClaw that mitigates the risks associated with running the agent on production hardware.
It’s crucial to conduct a 30-day stress test before expanding access, using throwaway identities and monitoring how the agent handles various tasks without exposing sensitive data. Testing in a sandbox environment allows for adversarial tests that would be too risky on production hardware, ensuring that the agent behaves as expected.
By implementing an evaluation framework that includes isolated execution, tiered integrations, and structured validation, you can mitigate the risks associated with agentic AI deployments. Building this infrastructure now will help your organization stay ahead of the curve and avoid potential breaches caused by AI agents like OpenClaw.
