New AI agent vulnerability may allow cross-model contagion, report warns

Agents in large language models such as ChatGPT, Claude and Gemini now operate autonomously and share files across sessions. According to Hipertextual, this interconnection has reportedly created a vulnerability that could spread between agents.

The flaw may let malicious files act as "mental viruses," compromising agent memory and behavior. Researchers note that while agents collaborate and retain session data, this same persistence could be exploited to alter their operations.

Hipertextual reported that the issue highlights risks in multi-agent ecosystems where models interact and share information.