Palo Alto Networks has launched Unit 42 Continuous Frontier AI Defense, an agentic offensive security service that uses gated cyber models, including Anthropic's Claude Mythos and OpenAI's GPT, to keep finding, validating and fixing enterprise exposures, the company said on 22 September 2026.
The service starts with a full-estate baseline scan and then tests continuously as an organisation's environment changes. A proprietary multi-model harness routes each task to the model best suited to it, using Claude Mythos 5, GPT-5.6-Cyber and open-weight models, and Unit 42 experts verify findings and chain end-to-end attack paths across web apps, APIs, cloud infrastructure, source code repositories and network assets.
It builds on Unit 42's point-in-time Frontier AI Defense, launched in April. Palo Alto Networks said it developed the approach over six months of in-house testing and more than 100 Unit 42 customer engagements, backed by a $17 million investment. In customer assessments, the Frontier AI Exposure Analysis found exposures in 100% of customers, and 37% of those were rated high or critical.
“AI has created an asymmetric advantage for threat actors against organizations trying to defend at human speed. Modern cybersecurity requires machine-speed defense,” said Sam Rubin, Senior Vice President of Unit 42 at Palo Alto Networks.
The service is available worldwide on an annual subscription, with options that vary by the OpenAI, Anthropic and open-source models used. These efficacy figures are vendor-reported.