Cato Networks evolved secure access service edge (SASE) to continuously assess AI agents in real time.
Declaring zero trust insufficient in the AI era, the firm explained that its Cato AgentSense tool traces how instructions and data influence an AI agent’s behavior at runtime, as well as flagging any deviations from its original mandate.
The service blocks any deviant actions, providing protection against indirect prompt injection risks. According to Cato, the threat is a real one, pointing to recent attack vectors like EchoLeak, a zero-click vulnerability in Microsoft 365 Copilot that allowed attackers to stealthily steal sensitive organizational data without requiring any action from the victim. The SASE vendor also highlighted CurXecute, a remote code execution flaw disclosed in 2025 that targets the Cursor AI code editor through its model context protocol (MCP) integration, and DuneSlide, a set of flaws that allowed prompt-injected instructions to bypass Cursor's built-in sandbox protections and trust boundaries.
“These examples underscore why enterprises need to trust that an AI agent will carry out the task they authorized, even when it encounters content an attacker controls,” said Itay Ravia, head of Cato AI Labs. “AgentSense helps preserve that trust by tracing how untrusted content influences an agent’s decisions and sensitive actions, helping enterprises adopt AI agents without giving attackers control of the workflow.”
Cato AgentSense is integrated within Cato's flagship SASE platform and is included at no additional charge within Cato AI Usage licenses. The service is derived from Cato's $320 million acquisition of Aim Security, an AI security firm specializing in securing enterprise use of large language models (LLMs) and generative AI.
In conversation with this title, Cato CEO Shlomo Kramer argued today's security stacks need "to cover the prompts and responses and action of agents across the muscle of the organization. It has to build the whole lifecycle of security, from posture management to protection to detection response, and back to posture.”
“This is a completely new security stack that is very different than the network traffic and the application and data security stacks,” Kramer said.