The security risks of letting AI agents act on your behalf
A sponsored report explains how AI assistants with delegated permissions can be tricked by hostile emails, websites, or add-ons into running commands, leaking information, or deleting files. It notes that no software bug is required because an agent may treat untrusted text as an instruction. Gen Digital describes an Agent Trust Hub meant to check extensions, secure connections, and enforce controls while agents operate.
Gen Threat Labs showed a controlled test in which an AI agent, while reading email, was steered by hostile content into running system commands, sending data outside its environment, and deleting files. The agent needed no software flaw; it followed text it should not have trusted.
The company also reported Warden Stealer, malware aimed at AI assistant and coding-tool configuration and local data, which could expose credentials, tokens, and chat histories. Separately, it found 12% of ClawHub skills at the time were malicious, often posing as legitimate add-ons to push malware.
People and organizations delegating tasks to AI agents could face data leaks, unwanted commands, or lost files if hostile content or add-ons slip through. Developers and businesses adopting agent extensions may need stronger vetting and visibility into agent behavior, while ordinary users may bear risks from credentials and conversation histories stored locally. Trust tools and runtime controls may shape adoption, but their effectiveness could determine whether agent use expands safely.