Traditional firewalls can’t see what’s inside an AI prompt. Check Point built one that can.

Nearly half of organizations are entirely blind to the machine-to-machine traffic their AI agents generate, and the tools built to watch enterprise networks were never designed to look inside a prompt. Check Point’s AI Network Firewall, launched this summer, is one recent attempt to close that gap from inside the infrastructure enterprises already run.

According to one recent study, 48.9% of organizations have zero visibility into the machine-to-machine traffic their AI agents generate. This speaks to a specific visibility problem, not just anxiety about AI: those organizations have no way to monitor what their autonomous agents are doing across connected systems.

That matters even more as agents are increasingly used for purchasing access to data and online services, interacting with customers, and completing tasks without a person approving every step. It also changes what organizations need from their security monitoring. A log showing that one service contacted another cannot, on its own, explain whether the agent was following the company’s instructions.

One reason for this blind spot comes down to how cybersecurity solutions were traditionally built in the first place. Legacy web application firewalls and basic API gateways were built around signatures, rate limits and predictable human sessions. An agent can improvise a new sequence of otherwise legitimate requests without matching a known attack signature. These tools simply were not built to parse the unpredictable, yet logic-based actions an autonomous agent can generate.

An approved API call is not automatically an approved business decision. Consider a support agent that uses valid credentials to retrieve a customer file and then includes it in a response to someone who should not receive it, without performing any tasks that resemble a conventional intrusion.

Encryption creates a separate obstacle. AI traffic traveling over HTTPS is encrypted, so a firewall needs appropriate certificates and policies in place to decrypt that traffic before it can examine the contents. Capacity and connectivity constraints also influence which traffic gets inspected. A readable prompt and an understood prompt are not the same thing: decryption might expose the text, but it does not establish whether the instructions are safe.

That distinction points to a broader confusion worth clearing up: there are two very different things people mean by “AI firewall.” An AI-powered firewall uses machine learning to detect conventional network threats. A firewall built to protect AI inspects prompts and agent interactions for AI-specific harm. The gap here concerns the second category, since prompt injection can turn a document or webpage into instructions an agent follows, the same mechanism used in recent attacks against coding agents.

Announced this past July, Check Point’s AI Network Firewall targets that gap by adding AI-specific inspection to existing firewall infrastructure. The company says that the AI firewall discovers and classifies employee use of generative-AI tools, AI agent and Model Context Protocol activity, and traffic to and from AI applications. Model Context Protocol, or MCP, is the industry standard for connecting agents to tools and data.

The product then applies real-time inspection to that activity. Check Point says this can identify sensitive data heading toward a public AI tool or a manipulated prompt attempting to trigger unintended behavior. Its broader AI security stack incorporates technology from Lakera, the AI security startup Check Point acquired in 2025, which supplies runtime protection against prompt attacks.

What sets this approach apart is the deployment model. According to Check Point, customers can use their existing firewall infrastructure without adding new hardware or software, bringing AI controls into their established management environment. That means security teams can start governing AI traffic without first having to deploy and maintain a separate system.

Nightfall AI’s Firewall for AI takes a different route. The company describes its standalone offering as a “client wrapper” around generative-AI interactions, using APIs and software development kits to inspect content before it reaches a model. It scans for personally identifiable information, payment-card details, health information and secrets, allowing sensitive material to be removed before an application forwards a prompt.

Nightfall separately offers prompt-injection protection and conversational guardrails as well. These checks address conversation content and signals such as model-response refusals, rather than sensitive-data categories alone. Finding a payment card number and recognizing an attempt to redirect a model are different security tasks.

Neither company’s approach is necessarily more advanced than the other. Apparently, Check Point’s take is that AI-specific protection belongs inside infrastructure a company already runs, whereas Nightfall’s is that AI interactions warrant a dedicated layer within application workflows. For companies that need to protect their AI systems, the practical questions concern coverage, intervention and policy enforcement, not which product approach seems freshest.

Both approaches point to a wider focus on reliable AI infrastructure rather than model performance alone. An integrated control may fit established operations, while an application-level wrapper gives developers a specific point at which to filter model-bound data. Neither placement, by itself, guarantees that every relevant interaction will be inspected.

A business that cannot observe its agents’ interactions cannot confidently assess whether those agents are staying within their remit. Whichever architecture gains ground, the meaningful advance will be tooling that connects an instruction to an action and applies policy before harm occurs. What matters is whether security tools can see what AI systems are actually doing, and govern it, rather than just logging the traffic after the fact.