Palo Alto Networks bets on machine-speed defense with Continuous Frontier AI Defense

Unit 42, the threat research arm of Palo Alto Networks, launched Continuous Frontier AI Defense, a subscription service that moves security testing from one-off assessments to always-on machine-speed probing. The service uses restricted-access AI models from Anthropic and OpenAI to continuously attack customer environments, prove exploitability, and suggest fixes.

What Continuous Frontier AI Defense does

The service combines Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6-Cyber with open-weight models through a proprietary multi-model harness. Unit 42's software decides which model handles each task, routing work to the most capable model to balance results and cost.

After an initial estate-wide scan, the system continuously attacks web apps, APIs, cloud infrastructure, code repositories, and networks as they change. It then attempts to prove each weakness can be exploited, maps potential attacker paths, and recommends code fixes or temporary virtual patches.

The service builds on a one-off assessment Unit 42 introduced in April and expanded in August to run on the two cyber models. According to Palo Alto Networks, that earlier assessment found exposures at every customer tested, with 37% rated high or critical severity.

The multi-model harness is the engineering story

For builders evaluating AI security tooling, the interesting part is the multi-model harness. Unit 42 doesn't just let one model run everything. It allocates tasks to the most suitable model, which is a practical approach to controlling cost and improving accuracy. Anthropic claims Mythos has found more than 10,000 high-severity vulnerabilities in widely used software. OpenAI's GPT-5.6-Cyber is similarly gated. Using both models plus open-weight alternatives through a single harness lets Unit 42 optimize for different types of security testing without overpaying for premium model calls on simple tasks.

The company says it spent six months and $17 million developing the approach across more than 100 customer engagements. Internal testing on its own systems allegedly surfaced as many exposures in three weeks as a typical year of assessments.

Pricing and availability

Pricing is annual and varies based on the mix of Anthropic, OpenAI, and open-source models a customer selects. The service is available worldwide. Exact pricing tiers are not publicly detailed.

Caveats worth noting

The performance data comes from Palo Alto Networks and Anthropic, not from independent evaluation. The claim that AI-driven attacks reduce flaw-to-exploit time by up to 97% is also a vendor assertion. Builders should expect real-world results to vary by environment, model access, and the skill of the testing harness. Additionally, Anthropic briefly suspended access to its Mythos models in June to comply with US export controls, which highlights a dependency risk for any service built on gated models.