Summary

Researchers have tricked six widely used AI agent frameworks into running tool calls written by an attacker, simply by asking the underlying model to repeat the call back.

AI summaries can be wrong sometimes—always verify important details using the source article.