China’s National Computer Virus Emergency Response Center (CVERC) and the National Engineering Laboratory for Computer Virus Prevention Technology have warned about a supply-chain campaign aimed at AI agents. They found several counterfeit skill packages in repositories for the open-source OpenClaw agent framework, known in China by the nickname “lobster”. Each package contained malicious code. The warning was published on April 22 and widely republished by Chinese state media.
What happened
Skills are plug-in modules that give an agent new abilities. Analysts at the national collaborative virus analysis platform found that some packages in OpenClaw skill repositories copied legitimate ones but carried hidden logic. After installation, when the agent calls the skill, it quietly downloads and installs a trojan in the background. That malware steals private data, which CVERC says can lead to leaks of sensitive information and financial loss. CVERC listed the poisoned packages and their file hashes in a table and a downloadable appendix.
Why it matters
CVERC said directly that agent skill packages have become a new channel for spreading malware. This repeats an earlier warning from CNCERT in March, which listed plug-in poisoning as one of four main risk areas for OpenClaw, along with prompt injection, operator error and software flaws. Agents often run with wide file-system, network and credential access. A single poisoned skill can therefore expose much more than a typical malicious browser extension can.
Who is affected
Individuals and organisations that run OpenClaw agents and install community skills, especially from unofficial repositories. Agents that run next to production systems or personal-data stores carry the highest risk.
Recommendations
- Audit every skill before installing it. Review its prompts, external URLs, scripts and the tools it calls, and confirm they match the stated function.
- Isolate agent hosts from production systems and personal-data systems, or apply strict access controls so a compromised agent cannot be used as a stepping stone.
- Compare installed skills against the hashes in CVERC’s published IOC appendix, and submit suspicious files to the national analysis platform at
virus.cverc.org.cn. - Prefer pinned, vendor-verified skill sources, and keep an inventory of the skills each agent has installed.