New OpenClaw AI agent found unsafe for use | Kaspersky official blog
Source Kaspersky Security Blog
Description
We explore whether OpenClaw can be safely installed and configured, and the risks involved in running this experiment.
Impact
For European organizations, the OpenClaw vulnerabilities pose significant risks, particularly for those integrating AI automation into their IT environments or using Apple hardware extensively. The ability for attackers to remotely execute code with full system privileges can lead to complete system compromise, data theft, and operational disruption. Sensitive corporate data, private keys, API tokens, and communication histories can be exfiltrated, potentially leading to intellectual property loss, financial fraud, and reputational damage. The unmoderated plugin ecosystem increases the risk of supply chain attacks, where malicious code is distributed under the guise of legitimate functionality. Organizations experimenting with AI agents without proper isolation risk exposing critical infrastructure and user data. The high resource consumption of OpenClaw may also impact operational costs and system performance. Given the agent’s design flaws, any deployment in production or on devices connected to corporate networks could facilitate lateral movement and persistent threats. The lack of authentication and reliance on default trust assumptions make even less sophisticated attackers capable of exploiting these vulnerabilities. Prompt injection attacks further complicate defense, as they exploit inherent weaknesses in large language models, potentially bypassing traditional security controls. Overall, the threat undermines trust in AI automation tools and highlights the need for rigorous security practices in AI deployments.
Defensive guidance
1. Deploy OpenClaw only on isolated, dedicated hardware or virtual private servers (VPS) that are not connected to critical networks or primary user devices. 2. Implement strict network segmentation and firewall rules using an allowlist approach to restrict all inbound and outbound traffic to only necessary ports and IP addresses. 3. Avoid exposing OpenClaw administrative interfaces to the public internet; if reverse proxies are used, ensure they are correctly configured to prevent forwarding external requests as localhost traffic. 4. Use burner or isolated accounts for all messaging apps and services connected to OpenClaw to limit the impact of credential theft. 5. Regularly perform deep security audits using OpenClaw’s built-in audit tools and third-party vulnerability scanners to detect misconfigurations and vulnerabilities. 6. Prefer AI models with better prompt injection resistance, such as Claude Opus 4.5, to reduce the risk of malicious prompt exploitation. 7. Avoid installing unverified or unmoderated skills/plugins from public repositories; only use vetted and signed extensions. 8. Educate users about the risks of social engineering and the ClickFix technique to prevent inadvertent installation of malicious code. 9. Monitor network traffic and system logs for unusual activity indicative of exploitation attempts or data exfiltration. 10. Keep abreast of updates from the OpenClaw community and security researchers, applying patches and configuration improvements promptly once available. 11. Consider alternative AI automation solutions with stronger security postures for production environments.