The argument for a native cloud firewall is easy to understand. It is already in the platform, the billing is familiar, deployment is straightforward, and the operational team does not have to introduce another vendor. Those are real advantages.

Unfortunately, they are not evidence of threat-prevention effectiveness.

The NSS Labs 2026 Cloud Network Firewall test produced a result that is difficult to explain away as a marginal difference between competing products. Four specialist security vendors—Fortinet, HPE Juniper, Palo Alto Networks, and Versa—achieved Security Effectiveness ratings between 99.74% and 99.91%. Check Point scored still higher, at 99.95%, although its cost per protected Mbps placed it in the Neutral rather than Recommended category.

The three native cloud-provider firewalls scored 77.45% for Google Cloud and 0% for both AWS and Microsoft Azure.

That is a significant, and worrying, gap. It is also more complicated than saying security vendors are good and cloud providers are bad. Cisco, a specialist vendor, scored only 66.10%. Google substantially outperformed it. Azure blocked every malware sample on the paths it inspected yet still finished at zero overall. Strangest of all, one cloud provider scored 0%, yet is using one of the top performing products in its network. To understand the results, we need to look below the headline score.

What the overall score actually means

Security Effectiveness is not simply the percentage of attacks blocked. The 2026 calculation combines routing and access control, TLS functionality, stability, malware blocking, exploit blocking, and resistance to exploit evasions. Routing, TLS, and stability act as multipliers, so a failure in a foundational capability can reduce the overall result to zero.

This matters because the native products did not all fail in the same way.

  • AWS recorded 0% for both malware and exploit blocking. It therefore had no prevention effectiveness to preserve.
  • Azure blocked 100% of the malware samples and 27.31% of the exploits, but it did not inspect TLS in the tested configuration. Since encrypted traffic flowed without inspection, its TLS factor was zero, leaving its overall Security Effectiveness at zero.
  • Google blocked 100% of the malware samples and 99.82% of the exploits. Its weakness was evasion resistance: 55%, which reduced overall Security Effectiveness to 77.45%.

Those are three quite different engineering failures. A single league-table number tells us who finished where. The component results tell us why.

Access control is not threat prevention

Every product passed the routing, segmentation, and access-control tests. This is a pretty low bar, of course, since that was the original function of a firewall. The field separated when the test moved beyond deciding which connections were permitted and asked what happened inside those connections.

That distinction is central to the native-versus-specialist question. Cloud providers begin with the network fabric: routing, segmentation, elasticity, platform integration, service availability, and metered consumption. Security vendors begin with hostile traffic: exploit research, malware analysis, signature quality, protocol normalization, evasion resistance, and the operational consequences of false positives.

That is an interpretation of the evidence, not a laboratory finding about corporate priorities. But the results are consistent with it: basic policy enforcement was common; mature threat prevention was not.

There is nothing inherently wrong with a product designed primarily as a cloud-network control. The problem begins when buyers assume that the word “firewall” means the same depth of protection regardless of architecture, inspection capability, or research investment. It does not.

Detection content is a specialist discipline

The earlier tests provide useful context. In the Q4 2024 native-firewall mini-test, AWS Network Firewall used open-source Suricata rules. The report found more than 21,500 preconfigured rules but estimated that roughly 17,500 monitored outbound connections—activity after infection—while only about 4,000 focused on inbound traffic. Many rules lacked CVE identifiers, making coverage against known vulnerabilities difficult to establish. AWS blocked two of the 522 exploits tested.

Google’s service followed a different path. Its threat detection and prevention used Palo Alto Networks technology. Under the tested default configuration, it blocked 50.57% of the same exploit set. After Google applied its recommended settings, the result rose to 86.97%.

The contrast does not prove that partner-supplied content is always superior or that open-source rules cannot be effective. It supports a narrower, more defensible conclusion: high-quality threat prevention requires sustained investment in vulnerability research, content development, validation, deployment policy, and false-positive control. A ruleset does not become enterprise-grade merely by becoming large. Twenty thousand shallow or misaligned rules do not equal comprehensive coverage.

Nor is content development a one-off exercise. New exploits and malware arrive continuously. Existing detections must be checked against new evasions. Signatures that are theoretically accurate still have to survive production traffic without blocking legitimate business applications. That is a specialist engineering discipline, and the strongest 2026 results show what mature execution looks like.

Evasion resistance separates signatures from inspection

Google’s result is particularly revealing. Its baseline protection was excellent: 100% malware blocking and 99.82% exploit blocking. Presented canonically, almost every threat was detected. But exploit-evasion resistance was only 55%, reducing the combined score to 77.45%.

An evasion test asks whether a firewall sees the same traffic as the protected system after packets have been fragmented, reordered, segmented, encoded, or manipulated at different protocol layers. The signature may be perfectly capable of recognizing the exploit; it is useless if the inspection engine reconstructs the traffic incorrectly.

This is where years of adversarial testing tend to show. Mature products do not merely accumulate signatures. They develop protocol parsers, normalization logic, state handling, target awareness, and regression testing for the awkward cases attackers deliberately create. It is unglamorous engineering, but it is often the difference between a security product and a security feature.

The 2025 test showed the same problem even more starkly. Google blocked 96.60% of baseline exploits but failed critical evasion tests, producing 0% overall Security Effectiveness under that year’s scoring. In 2026 it improved substantially, but the remaining 55% evasion result shows that high baseline detection and resilient inspection are still not the same thing.

Encryption is an architectural gate

Azure illustrates another failure mode. The product blocked every malware sample presented to it but did not inspect TLS in the tested configuration. Traffic using the unsupported encryption path continued to flow without inspection.

It would be misleading to summarize that as “Azure detected nothing.” It demonstrably detected malware on traffic it could inspect. The problem is architectural: detection quality on visible traffic cannot compensate for an inability to inspect the traffic carrying most modern applications.

The test weighted rated throughput 95% toward TLS and 5% toward HTTP. Azure’s reported throughput was therefore based on HTTP alone and was explicitly marked as not comparable with the other products. That is the correct treatment. A firewall that is fast because it is not performing the difficult work is merely failing fast.

It also gets expensive fast. Azure recorded the lowest rated throughput in the test, at 216 Mbps, and the highest total cost per protected Mbps, at $129.80. Google cost $30.45 per protected Mbps and AWS $21.78. Fortinet, HPE Juniper, Palo Alto Networks, and Versa—all with Security Effectiveness above 99.7%—ranged from $12.48 to $17.60 per protected Mbps.

Cloud-native does not automatically mean cheaper once protection and throughput are included in the calculation.

Security specialists are not infallible

The specialist results prevent this becoming a lazy category argument.

Cisco blocked 96.44% of exploits and resisted 99% of exploit evasions but blocked only 36.73% of the malware samples. The addition of malware to the methodology reduced its overall Security Effectiveness to 66.10%.

Check Point produced the highest Security Effectiveness in the test at 99.95%, including 100% malware blocking, 99.91% exploit blocking, and 100% evasion resistance. Yet it received a Neutral rating because its $32.52 cost per protected Mbps exceeded the group average. Many customers might consider it is worth paying a little extra for such levels of protection, however.

Those results are useful precisely because they resist simplistic branding. A security pedigree does not guarantee balanced protection. A very effective product is not automatically an economical one. And a native product is not condemned to poor detection—Google’s baseline threat results prove otherwise.

The defensible conclusion is this: in the 2026 test, the mature specialist products generally delivered much stronger and more complete protection than the native cloud firewalls. Vendor category explains part of the pattern, but component testing explains the individual results.

A better procurement question

Enterprises should not ask whether native or third-party firewalls are universally better. They should ask what job the control must perform.

If the requirement is segmentation, cloud integration, and basic policy enforcement, a native firewall may be entirely adequate. If the control is expected to inspect encrypted traffic, prevent exploits, block malware, and resist deliberate evasions, each of those capabilities must be evaluated explicitly. Where a native service falls short, the answer may be a specialist firewall, a layered architecture, or a consciously limited role for the native control.

Convenience has value. So do consolidated billing and operational familiarity. But none of them blocks an exploit, detects a malicious file, or reconstructs an evasive TCP stream.

Security effectiveness remains a separate purchasing criterion. The test data says it deserves to be treated as one.