Facts
The open-source project OpenClaw is a locally running AI agent framework. It enables the installation of a partially autonomously acting AI agent that executes instructions and, depending on configuration and integration, can receive far-reaching system rights as well as access to keys and passwords for various services, AI models, websites, credit cards, social networks, and so on.
OpenClaw published an advisory on 31 January 2026 on a severe vulnerability (CVE-2026-25253) that enables remote code execution. Operators of OpenClaw are recommended to keep the software up to date in order to close vulnerabilities as quickly as possible, as well as to restrict reachability from the Internet. On 16 February 2026, the security company Hudson Rock reported that configuration files of the AI agent OpenClaw had for the first time been stolen by infostealer malware. This was not a targeted compromise of OpenClaw itself, but a malware infection on the affected systems. The malware automatically searched compromised systems for files with terms such as "token" or "private key" and stole, among other things, the files openclaw.json, device.json, as well as persistent storage files. In this way, authentication tokens and cryptographic keys could be intercepted, which potentially enables a complete digital identity takeover.
Assessment
AI agents are increasingly becoming a target of attacks because they bundle sensitive credentials and comprehensive context information. AI agents and frameworks are a still relatively new and untested technology that, through their deep system integration and extensive access rights to local data, pose a high security risk. They are often inadequately tested, partly faulty software.
Reaction
Due to vulnerabilities, conflicts with security rules and security concepts, potential data outflow, and data protection violations, OpenClaw should only be used experimentally and not on production systems and production IT networks. Organizations should set clear rules for their personnel and, if necessary, prohibit the use of OpenClaw. In addition, dangers arise from the use of OpenClaw as unapproved "shadow IT."
If OpenClaw is nevertheless used after a risk assessment, at least current updates must be installed and the manufacturer's best practices followed. In addition, the following technical security measures should be considered: use of isolated environments (e.g., containers, virtual machines, or sandboxes), system and API accesses should be granted only to the extent required for the task, secure management of sensitive credentials instead of storage in plaintext, restriction of network accesses and external API calls, activation of logging and monitoring to detect unusual agent activities, and regular security reviews of the workflows and plugins used.