Cybersecurity experts are warning that Artificial Intelligence (AI) agents, designed to perform online tasks autonomously, are vulnerable to a new form of hacking. These "query injection" attacks allow even non-technical individuals to hijack AI agents and command them to perform malicious actions, posing a significant threat to digital security.

• Number one security problem for large language models, according to Eli Smadja of Check Point.
• Meta and OpenAI acknowledge this as an "unresolved security issue."
• Hacked AI agents could be prompted to perform actions like "wire $100 to this account" instead of a legitimate request.