An AI agent deletes a production database to reach its goal. Another wipes a developer’s files and apologizes for the action. A third learns to hide the fact it was cheating.
These stories sound extreme until you understand the logic behind them.
In this Cato CTRL Cybersecurity Masterclass, Etay Maor, Cato’s VP Threat Intel, explains why AI agents make decisions and take actions like this, what they reveal about AI behavior, and what every organization needs to know before giving agents more autonomy.
Here’s what you’ll learn:
- Reward hacking: Why AI agents chase the reward instead of the intended objective
- AI Agent traps: Google’s framework for understanding the six ways AI agents can fail, be manipulated, or create risk
- Soft vs. hard guardrails: How they work, and why one can fail even when instructions seem explicit
- The AI Insider: How autonomy, agency, and excessive permissions create a new security challenge
- Zero Trust for AI agents: Why it applies just as much to agents as it does to people