Palo Alto Networks launched a subscription cybersecurity service Tuesday that places several advanced-artificial-intelligence models inside a continuously running offensive-security program, a move that tests whether defenders can use the same accelerating technology that is making automated attacks more capable.

The service, Unit 42 Continuous Frontier AI Defense, combines Anthropic’s Claude Mythos 5, OpenAI’s GPT-5.6-Cyber and open-weight models with Palo Alto’s security researchers. The company says it will repeatedly examine web applications, application programming interfaces, cloud systems, source-code repositories and network assets as customer environments change. The launch and core design were independently reported by Reuters and Axios, while Palo Alto published a formal company announcement.

Why the multi-model design matters

Traditional vulnerability scanning generally identifies known weaknesses or suspicious configurations at a particular moment. Palo Alto’s proposed system goes further: it is designed to test whether separate flaws can be chained into a workable attack path, then send prioritized remediation guidance into a customer’s ticketing process. The company’s product page says human offensive-security experts verify model findings before they are treated as exploitable paths and that testing restarts when applications or infrastructure change.

That distinction matters because a list of isolated defects does not show how an intrusion would unfold. A low-severity configuration error can become important when combined with exposed credentials, excessive cloud permissions or an unpatched service. Conversely, a scanner may flag a weakness that compensating controls make unreachable. The advertised workflow attempts to move from detection to proof: map possible routes, validate which ones work and direct remediation toward the paths that create the greatest risk.

That human review is important because the company is not claiming that one model can reliably see an entire environment. Palo Alto told Axios that, in its own testing, no single model found more than 40% of the vulnerabilities in a complex customer environment. It also said the findings from Mythos 5 and GPT-5.6-Cyber overlapped less than 10% of the time. Those figures are vendor-reported rather than results from an independently published benchmark, so they should be read as evidence for the company’s architecture—not as a universal measure of model performance.

The underlying argument is that different models have different blind spots. One may be stronger at reading source code, another at reasoning across cloud permissions, and another at linking a configuration error to an exposed service. An orchestration layer can assign tasks to different systems and compare their results. Human specialists then decide whether a theoretical weakness is reachable, consequential and safe to test.

From periodic testing to continuous pressure

The shift is operational as much as technical. A conventional penetration test may produce a valuable report but can become stale after software releases, access changes or cloud deployments. Palo Alto says its service continuously discovers exposures, validates exploitability and re-tests after changes. Reuters reported that customers can receive code-level repair suggestions or use virtual patches when an official software fix is not yet available.

A virtual patch can reduce exposure by blocking a malicious request or attack pattern at a security layer, but it does not remove the defective code underneath. That makes follow-through measurable: organizations still need to assign owners, test permanent fixes and confirm through re-testing that the attack path is closed. A continuous service could shorten that loop, although the launch materials do not provide independently verified remediation-time improvements.

Annual subscriptions will be offered globally, with pricing based partly on the mix of commercial and open-weight models selected, according to Reuters and the company. Palo Alto has not publicly provided enough detail to compare the service’s cost with periodic penetration testing or competing platforms. It also has not released customer-level false-positive rates, benchmark methodology or a complete account of how model access and sensitive enterprise data will be governed.

Those unanswered questions matter. Offensive testing requires powerful access, carefully defined scope and strong audit controls. An automated system that misidentifies a target, executes an unsafe action or exposes proprietary code could create new risk while searching for existing risk. Palo Alto says scope is agreed with the customer and expert validation is built into the workflow, but prospective buyers will still need contractual clarity on data retention, model-provider access, incident responsibility and the boundaries of autonomous action.

A governance test as well as a security product

The launch arrives as federal guidance increasingly treats AI deployment as a risk-management problem rather than a one-time technical purchase. The voluntary NIST framework calls for trustworthiness considerations throughout the design, use and evaluation of AI systems. NIST also describes its cyber framework as a structure for governing, identifying, protecting, detecting, responding to and recovering from cybersecurity risk. Those documents do not endorse this product; they illustrate the wider controls against which any continuous AI security service should be assessed.

For security leaders, the practical question is not whether an AI model can find a flaw in a demonstration. It is whether a governed combination of models and people can repeatedly find meaningful exposures faster than adversaries, without generating unacceptable operational or data risk. Palo Alto’s new service turns that question into a commercial offering. Its value will depend on independent evidence, transparent controls and whether customers can convert a faster stream of findings into verified fixes.