AI Agents: "Almost Infinite" Security Risk, CEO Warns

An AI security CEO warns that the risk from AI agents is "almost infinite." Noam Schwartz, CEO of Alice, explains that companies spend months testing AI models for safety. However, when these models connect to tools like Slack, Gmail, and databases, the initial safety testing becomes irrelevant. The moment AI agents can take actions, the number of potential problems explodes. Schwartz notes that agent risk increases dramatically when AI can act, rather than just produce text. He describes how AI can automate tasks that once required large criminal organizations. Attackers only need one successful path, while defenders must cover everything. Security needs to be present at every layer, including models, tools, data, and permissions. Agents can even influence other agents, a process likened to radicalization. The bottom line is that AI security extends beyond traditional cybersecurity to include fraud detection, threat intelligence, and trust and safety. This matters because it offers a practical definition of "safe" for anyone building or deploying AI agents.