AI agents and malicious plugins raise security concerns

AI agents that operate autonomously—downloading files, running scripts and even making payments—pose a far greater security risk than traditional generative chatbots such as ChatGPT, Claude or Gemini, which only react to user prompts. These agents rely on AI skills (plugins) that extend their capabilities, but the skills can contain hidden malware. A March 2026 ESET analysis of 800,000 AI skills identified 25,000 as suspicious and more than 2,500 as outright malicious, including infostealers and cryptomining tools. The marketplace ClawHub alone listed over 800 harmful skills. Security experts advise treating AI skills like software: verify required permissions, prefer skills from known, trusted developers, and monitor what agents are allowed to do.

The guidance emphasizes that while chatbots have become safer, the real danger now lies with AI agents and their plugins, which can act without explicit user consent and expose devices to attacks.