Tal Peretz’s Post

Every time Travis McPeak and I are in the same room, the hot takes keep flying. Here are 6 spicy truths about agent security you won't hear at RSA: 1. Nobody is using just one AI client. Even people at Cursor use Claude, Codex, and whatever shipped last Tuesday. Your security model has to assume multi-client, multi-agent, all the time. 2. Visibility without action is a dashboard, not security. If you can't trace root cause and enforce controls in real-time, you're just watching the breach happen in HD. 3. The real risk isn't code leaking - it's customer data leaking. The moment you connect agents to external tools like Snowflake, Salesforce, or Slack, you've created an attack surface no code review catches. 4. Non-technical people are already using AI tools across every department. You can't train your way out of that. The only answer is secure by design - make the unsafe path impossible, not just discouraged. 5. AI is better than most engineers at securing code. As code generation accelerates, the only viable play for security teams is fighting agents with agents. 6. With all that risk, here's the biggest one - being the company that doesn't enable agents fast enough. Slow adoption isn't caution, it's a different kind of risk. Travis is building the security layer at Cursor. We built the golden path and control plane for agents 🤝