OpenClaw’s Unreviewed Code and Malicious Skills Raise Security Risks
Summary
OpenClaw can handle sensitive tasks but may expose users to risks from largely unreviewed code, malicious skills, and prompt injection. The article advises testing it in an isolated system without sensitive data or transaction access.
Key points
- OpenClaw can connect multiple actions and may be given access to calendars, payment details, accounts, email, and messaging.
- Its developer reportedly has not reviewed most of the code, which the article says may leave security vulnerabilities undiscovered.
- The article says a frequently downloaded skill is effectively an infostealer that can send users’ credentials, payment information, and access tokens to attackers.
- Malicious skills may contain malware that OpenClaw downloads and executes; prompt injection can also hide instructions for the agent on webpages.
- OpenClaw has begun removing malicious skills and, according to its official website, is cooperating with VirusTotal.
- The article recommends experimenting only in an isolated system, inaccessible from outside, and keeping data needed for payments or other binding transactions off it.
Article Details
- Topic
- Security risks of OpenClaw agentic AI, including malicious skills, prompt injection, and access to sensitive data