Product Demo Contact us
eBook

The Lethal Trifecta

Keep your AI agents working for you, not an adversary

When an AI agent can access private data, read untrusted content, and communicate externally, the “lethal trifecta” is created. Attackers exploit that combination through indirect prompt injection, planting malicious instructions in content the agent reads.

The agent can carry out those instructions within its legitimate permissions, allowing an attacker to steal sensitive data without direct system access. Removing a capability can reduce exposure, but it can limit what the agent was deployed to do. So, how do organizations preserve critical capabilities while keeping agents safe?

This eBook explains the risks of the lethal trifecta, why security controls can miss the manipulation, and how Cato AgentSense helps enterprises trust the actions their agents take.

What’s inside:

  • The lethal trifecta: Three capabilities that make an agent useful, and give attackers a path to data theft
  • The permissions gap: Why valid permissions don’t establish whose instructions an agent is following
  • What Cato AI Labs uncovered: How researchers used prompt injection to demonstrate data theft in Microsoft 365 Copilot and a path to system compromise in Cursor.
  • The limits of existing controls: What risks remain when you restrict agents’ access to data, untrusted content, or external communication

Download the eBook