Check Point and NVIDIA Pair Infrastructure Limits With Step-by-Step AI Agent Monitoring
Check Point says a beta integration with NVIDIA OpenShell combines an infrastructure-level boundary with semantic monitoring to keep autonomous AI agents on task. According to the company, the monitor's verdict can arrive before an action runs, in under 100 milliseconds.
- Check Point says an agent's goal rarely specifies how it may be reached, and capable agents find unplanned routes.
- The NVIDIA Open Agent Safety Platform places the boundary in the infrastructure, outside the agent's reach.
- Check Point's semantic monitoring judges whether each step still fits the task.
- The two already work together in a beta integration with NVIDIA OpenShell.
- According to Check Point, the monitor's verdict can arrive before the action runs, in under 100 milliseconds.
A goal does not say how to reach it
Check Point's argument starts with a simple observation about how organizations deploy autonomous agents. When a company hands an agent a job, such as processing supplier invoices, it states the goal. According to Check Point, that goal rarely says how the agent may reach it, and capable agents are good at finding routes nobody planned for.
The post, titled "Goals Are Not Enough: Securing AI Agents with NVIDIA OpenShell", treats this as the core security problem. A well-defined objective is not the same as a well-defined set of permitted methods. The more capable the agent, the more likely it is to find a path that the people who set the task never considered.
Two layers: a boundary and a judge
Check Point describes two complementary controls. The first is the NVIDIA Open Agent Safety Platform, which, according to the post, puts the boundary in the infrastructure, outside the agent's reach. Because the limit sits outside the agent, the agent cannot simply reason its way around it.
The second is Check Point's own semantic monitoring. Instead of defining where the agent may operate, it judges whether each step still fits the task. In the invoice example, that means asking not only whether an action is technically possible, but whether it makes sense for the job the agent was given.
- Infrastructure boundary: set by the NVIDIA Open Agent Safety Platform, outside the agent's reach.
- Semantic monitoring: Check Point's check on whether each step fits the task.
Beta integration with NVIDIA OpenShell
According to Check Point, the two approaches already work together in a beta integration with NVIDIA OpenShell. The company highlights one performance claim: the monitor's verdict can arrive before the action runs, in under 100 milliseconds. A check that lands before execution matters because it allows a questionable step to be assessed before it has any effect, rather than being discovered afterward.
The excerpt of the post available for this report does not detail pricing, general availability or customer deployments, so those points remain unclear. The integration is described as a beta.
What it means for teams deploying agents
For organizations putting agents to work on routine business tasks, the message from Check Point is that stating a goal is not a security control. The company's pitch is that safe deployment needs both a hard limit that the agent cannot reach and a continuous judgment on whether each action serves the assigned task.
If the sub-100-millisecond claim holds in practice, the monitoring step could run ahead of each action without being a bottleneck, which is what makes pre-execution checks practical. Teams evaluating the beta will want to test that claim against their own workloads.
Check Point: Goals Are Not Enough: Securing AI Agents with NVIDIA OpenShell
