Cybersecurity

Researchers Demonstrate Prompt-Injection Attack Against Autonomous Agents

The technique hides instructions in ordinary web content, and the team says defences that rely on filtering alone did not hold up.