
Cato Networks offers $30,000 to hackers who can trick an AI agent
The Israeli cybersecurity company is inviting researchers to attack an AI workflow protected by its new AgentSense product. The first three eligible participants to successfully divert the agent from its assigned task will each receive $10,000.
Cybersecurity unicorn Cato Networks is offering a total of $30,000 to the first three eligible participants who successfully trick an AI agent into deviating from its assigned task by injecting malicious instructions. Each winner will receive $10,000.
The company, founded by Shlomo Kramer, is announcing Cato AgentSense today, a patented security solution developed by Cato AI Labs that continuously assesses the activity of autonomous AI agents in real time. Its purpose is to protect the integrity of agents' decisions, actions, and outputs, and to prevent them from carrying out harmful actions that conflict with their users' intentions.
As part of the challenge, Cato is inviting members of the cybersecurity community to attack an AI agent workflow protected by AgentSense. Participants will attempt to embed malicious instructions in content accessed by the agent, with the goal of diverting it from its assigned task and prompting it to perform unauthorized actions. Cato will use the findings and feedback to evaluate and strengthen the solution's defensive capabilities.
"In an era where organizations entrust AI agents with decisions that impact their revenue, customer relationships, and business continuity, overseeing the activity of these agents becomes just as critical as financial oversight," said Shlomo Kramer, CEO and co-founder of Cato Networks. "Granting greater autonomy to AI agents necessitates security mechanisms capable of thwarting harmful actions. Our vision is for AgentSense to provide our clients with the confidence to rely on AI agents, even for managing critical business processes, without losing control over them."
Even when AI agents operate with legitimate permissions, they can still take dangerous actions that fall outside the scope of their assigned tasks. Monitoring their access to systems and data is therefore not enough; organizations must also track how agents use the permissions they have been granted. Cato argues that traditional access controls and Zero Trust security measures must be supplemented by monitoring an agent's behavior while it is carrying out a task.
A key source of risk stems from a fundamental weakness in the architecture of AI agents: their difficulty in consistently distinguishing between legitimate instructions they are supposed to follow and external content they encounter while carrying out a task. Attackers can exploit this weakness by embedding malicious instructions in emails, web pages, documents, tickets, or responses generated by other tools, potentially diverting an agent from the task assigned by its user. This type of attack is known as indirect prompt injection.
When an agent has access to sensitive organizational data, is exposed to untrusted content, and can use external tools to take autonomous actions, it creates what is known as the "lethal trifecta." This combination can turn prompt-injection attacks into a serious business threat, potentially leading to data leaks, compromised systems, unauthorized actions, and disruptions to business processes.
AgentSense is designed to address this risk through two capabilities. The first, Untrusted Influence Detection, traces how instructions and data influence an agent's behavior while it is operating. When the system detects that untrusted content is steering the agent toward an unsafe action, it is designed to block that action before it is executed. Cato says this approach can detect and stop indirect prompt-injection attacks even when they do not match previously identified patterns, rather than relying primarily on static classifiers or known attack signatures.
The second capability, Drift Detection, is designed to identify when an agent's actions depart from the original task or user intent. It evaluates whether the steps an agent takes as a workflow progresses continue to serve the intended objective, including cases in which individual actions may be permitted but collectively fall outside the task it was authorized to perform. Cato describes this as a complementary layer of protection alongside Untrusted Influence Detection.














