- Cato Networks’ new AgentSense tool stops AI coding agents from leaking sensitive data or taking destructive actions
- Gartner says 71% of organizations will deploy agentic AI within 12 months, but only 21% have governance in place
- AgentSense runs on endpoints and supports Codex, Claude Code, Cursor and GitHub Copilot
Enterprises have handed AI agents the keys to the kingdom. Cato Networks wants to make sure they don’t hand them to a stranger. The company this week launched AgentSense, a new tool designed to stop valid permissions from turning into unsafe action.
“We're relying on the judgment of agents taking tasks autonomously, and those agents need guardrails in place,” Brian Anderson, Cato’s global field CTO, told Fierce. “When you mix together the interaction of sensitive and trusted sources, untrusted destinations and high risk operations and access to sensitive data, that is where the exposure can reside.”
What’s the problem?
A recent survey from Palo Alto Networks found nearly three-quarters (73%) of organizations let AI agents modify code and pipeline configurations and production environments. And a Q2 2026 Gartner survey of more than 2,000 CIOs and technology executives found 71% of organizations expect to deploy agentic AI within the next 12 months, yet only 21% currently have agentic AI governance capabilities in place.
“That imbalance makes it difficult to create value faster than AI creates risk. Managing value and risk is about building the right kinds of momentum while avoiding the wrong kinds of drag,” Gartner VP and Distinguished Analyst Kristin Moyer said in a statement.
The problem, Anderson told Fierce, is that enterprises are granting agents more autonomy and often excessive permissions but providing little in the way of guardrails on the judgments those agents exercise. And that’s a huge problem – one so big the World Economic Forum has begun issuing stark warnings about the need for enterprise AI guardrails.
Agents with too many permissions and not enough oversight can lead to unintended consequences. For example, an agent may be allowed to read corporate data in SharePoint or Slack. It may separately be allowed to access the web or other external tools to complete a task. But when the agent ships internal data off to an untrusted external destination? Well, that’s not exactly what its permissions were intended to enable, and that kind of action could lead to the exposure of proprietary or sensitive data.
There’s also the risk that an agent – in accessing either internal or external tools – encounters a new set of instructions hidden in an email, web page or support ticket. Remember how you used to pad term papers with white text to hit the required word count? Or how you embedded keyword tags in white font on a webpage to be ranked higher in search results? It’s like that, but with agentic prompts.
Anderson calls this task redirection “agent drift” and it can be dangerous if not spotted.
Keeping agents in line
AgentSense’s answer is to track where an agent’s instructions come from and where its output is headed.
Anderson said the product relies on definitions security teams already understand: trusted sources such as SharePoint, Outlook and Slack, and what counts as sensitive data. It then asks whether an instruction came directly from the user or from an outside influence, and whether the action would send sensitive data to an untrusted destination or perform a high-risk operation.
When those factors line up, it raises a red flag. “It’s kind of looking for those two key factors to align,” Anderson said.
All told, AgentSense is built to catch three categories of risky behavior: untrusted input to high-risk operations, in which instructions from an outside source, such as a web page or email, trigger a dangerous action; external exposure of sensitive information, in which sensitive data flows to a destination the organization doesn’t trust; and critical resource changes from untrusted input, such as deleting a database table or sending data somewhere it was never meant to go because of an outside instruction.
Today AgentSense works with coding agents the Cato platform already supports, including Codex, Claude Code, Cursor and GitHub Copilot. But Anderson said it’s looking to expand to different kinds of agents as well.
Asked why it focused on coding agents first, Anderson said it’s because coding agents have some of the highest adoption rates but come with the biggest risk.
“I give root keys, API keys get exposed, sensitive data that people have on their workstations are exposed, where they're mixing personal subscriptions with Claude with corporate subscriptions as well,” he explained. “So the highest risk that we've seen really starts with these coding agents.”
Bounty program
To put AgentSense through its paces, Cato has invited external security researchers to try to break it with a bounty program. The company is offering $10,000 each to the first three eligible participants who redirect a protected workflow.
Anderson said the goal is to validate the product’s efficacy and make sure it can handle whatever shows up in the wild, “battle testing” it as it hits the market.
Following the program’s launch earlier this week, thousands of people registered and participated, a Cato PR representative said. No one had cracked it within the first 24 hours.
Read more about AI governance on Fierce Network:
AI agents are joyriding through production environments
AWS wants to put AI agents on a Layer 7 leash
IBM reins in rogue AI agents with identity and runtime controls