Product

Palo Alto’s Unit 42 launches always-on frontier AI defense

The subscription runs Claude Mythos 5 and GPT-5.6-Cyber to find and help fix enterprise flaws, Palo Alto Networks says.

  1. Palo Alto Networks IR / PRNewswire
Glass-and-beige office towers of the Palo Alto Networks campus under a clear blue sky, small company mark on the entrance canopy, parked cars at right.
Palo Alto Networks headquarters. Photo: Namaste jinx / Wikimedia Commons (CC BY-SA 4.0)

Palo Alto Networks (NASDAQ: PANW) launched Unit 42 Continuous Frontier AI Defense — an always-on, agentic offensive-security subscription that runs gated cyber models (Anthropic Claude Mythos 5, OpenAI GPT-5.6-Cyber, plus open-weight models) through a proprietary multi-model harness to find, validate, and help remediate enterprise exposures. Worldwide on an annual subscription; SKUs vary by which OpenAI / Anthropic / open-source models the customer buys into.

What it does: Full-estate baseline scan, then continuous testing as the environment changes. Harness routes each security task to the model Unit 42 thinks fits. Advanced adversary simulation validates end-to-end attack paths across first- and third-party web apps, APIs, cloud, source repos, and network assets. Remediation package: prioritized fixes, code-level guidance, virtual-patch recommendations (pairable with Frontier Virtual Patching before public disclosure / official patches).

What it is not: A consumer product or a replacement for human IR. It is consulting-arm tooling + gated models — attackers with the same class of AI is the threat pitch Unit 42 is answering.

Lineage: April Frontier AI Defense (point-in-time Exposure Analysis + security blueprint). August: Exposure Analysis expanded with GPT-5.6-Cyber and Mythos 5. Sep 22 ship turns that into continuous / always-on service.

Company efficacy (hedge hard): Six months internal + 100+ customer engagements; $17M R&D claim. Internal: a year’s worth of exposures in three weeks. Customer Exposure Analysis: exposures in 100% of customers assessed; 37% high or critical; most from first-party apps; more than two in three third-party exposures had no known CVE.

Who for: Enterprises that want gated Mythos / GPT-Cyber capability wrapped in Unit 42 process — not DIY frontier-model red teaming. Replaces one-off Exposure Analysis with a subscription loop.

Risks: Vendor efficacy tables. Continuous offensive scanning can itself be risky if poorly scoped. Model access is gated for a reason — misuse and supply-chain trust sit with PANW’s harness and contracts. Price opacity until sales.

— Rowan Keel

Sources

  1. Palo Alto Networks IR / PRNewswire
  2. PR Newswire (same wire)

What’s next

  • Independent audits of Unit 42 efficacy claims.
  • Pricing disclosure beyond annual-subscription framing.
  • Scope incidents from continuous offensive scanning.

Edition Sept. 23, 2026 · Archive shelf