AI Agents Can Be Hijacked Without Being Hacked

Summary

The video explains how indirect prompt injection in emails or malicious tools can misuse an AI agent’s legitimate permissions, and describes monitoring and controls Radware says its AI security offering provides.

Key points

• An attacker could hide instructions in an ordinary-looking support email that an AI agent reads, even if a human employee does not notice them.
• If the agent treats analyzed content as instructions, it could retrieve customer records and send them to an attacker using its legitimate permissions.
• The video describes this as indirect prompt injection and notes that the scenario does not require malware or a stolen password.
• A malicious or compromised third-party tool could behave normally until a specific condition triggers unexpected actions.
• The video recommends visibility into agents, their tool access, the information they handle, and the actions they take.
• Radware says its Agnostic AI protection monitors risks including prompt injection, agent behavior hijacking, tool misuse, supply-chain attacks, and rogue agents; it also describes a security plugin with guardrails and access governance.

Vendors

Products