Check Point Adds Agent Security Layer to Google Cloud's Gemini Enterprise

Why authorized agents still need protection

According to Check Point, AI agents are moving into production quickly. They read enterprise data, call tools and take actions across business systems. That creates a problem: an agent can be fully authenticated and authorized and still be manipulated into doing the wrong thing.

Google Cloud's Gemini Enterprise offers enterprises a central control plane for agents. Check Point says it extends that control plane with AI-specific risk assessment and context-aware protection. In July, Google Cloud announced enhancements to Gemini Enterprise, including two features at the center of the integration:

• Agent Registry: a single view of all agents, servers and connections built across an organization, which lets administrators monitor agent sprawl.
• Agent Gateway: a central control point to secure and govern interactions across an organization's agent ecosystem. It allows granular access controls and safeguards against prompt injection, tool poisoning and data leakage.

Where Check Point plugs in

Check Point AI Agent Security integrates with both services, plus agent observability. The company describes three integration points.

• Pre-workflow risk posture (Agent Registry): Check Point observability reads the registry to surface risk posture on agents and their connected tools before workflows begin. Teams can see risky configurations, permissions and Model Context Protocol (MCP) connections and set boundaries before agents reach production.
• Prompt defense (Agent Gateway): all user-to-agent and Agent2Agent (A2A) traffic is checked by Check Point prompt injection defenses, embedded as Agent Gateway service extensions. Inspection happens inline, before the agent acts.
• MCP security (Agent Gateway): MCP tool calls are catalogued and permissions are checked against agent identity. Indirect prompt injection carried in MCP server content is blocked in real time.

Check Point says the key design choice is to evaluate not only whether an agent has permission to perform an action, but whether the action is appropriate given the full context of the interaction. Prompts, responses, tool calls and multi-step behavior are all in scope.

Three questions for security teams

Check Point organizes the integration around three questions. The first is where the agents are. Agent Registry provides a centralized view of agents, components, tools and MCP connections, and Check Point adds posture assessment and policy controls on top.

The second is whether production agents can be protected. In the company's example, an attacker uses prompt injection to steer an agent into calling an approved tool for an unintended purpose. Check Point, working in line with Agent Gateway, assesses the input, the tool response and the intended action together, and blocks the unsafe step before execution.

The third is whether sensitive data can be kept inside. Check Point says it inspects interactions at the gateway to identify sensitive data and how it is used, and enforces policy before information is returned, transferred or handed to a downstream tool.

Production figures and availability

Check Point says the runtime engine behind the integration already protects agent traffic at scale. At Dropbox, it secures millions of daily AI interactions with 99.8% of real-world threats blocked, sub-40ms average latency and a production false positive rate under 0.5%. At Nubank, it secures over 1M transactions per app per day across 100+ languages at a 0.01% false positive rate. According to the company, deployment took minutes, with zero code changes to customize policies.

The company argues that speed matters on a gateway, because security that adds latency or noise gets turned off. The offering is powered by the Check Point AI Defense Plane for employees, applications and agents. Customers on Google Cloud can discover their agent estate, assess risk, set policy and inspect agent behavior at runtime, while development teams keep building.