Palo Alto Networks to run OpenAI cyber models inside customer networks

Palo Alto Networks announced that its Unit 42 consulting team will integrate OpenAI Group PBC's advanced cyber models into customer networks. This expansion, called Unit 42 Frontier AI Exposure Analysis, utilizes the models to identify vulnerabilities, misconfigurations, leaked credentials and unmanaged attack surfaces across applications and network assets.

The models then undergo adversarial simulations to determine the exploitability of identified exposures and the potential attacker's progression once inside. Unit 42 discovered that 36% of the exposures identified have no known Common Vulnerabilities and Exposures (CVEs), many of which involve multiple interconnected gaps. The company emphasizes the importance of discovery and testing in uncovering these hidden vulnerabilities.

Unit 42 gained access to OpenAI's Daybreak program, specifically the Daybreak Red tier, which utilizes a purpose-built model called GPT-5.6-Cyber. This model performed exceptionally well in OpenAI's benchmark, correctly addressing 95% of advanced cybersecurity requests, compared to 1.5% for the general-purpose GPT-5.6 Sol. Palo Alto Networks was listed among the program's partners.

The multi-model harness allows Unit 42 to assign tasks to the most suitable model for each task, thereby broadening its coverage. Human review remains an integral part of the process, with Unit 42 consultants guiding the models and validating their results using Palo Alto Networks' telemetry and the company's own threat intelligence.

Remediation plans, ranked by the potential impact on attack paths, are then developed and integrated into existing IT, development, and security workflows. Since its launch, Frontier AI Exposure Analysis has been briefed to over 1,000 security teams, and hundreds of customers have been introduced to the service. This collaboration is part of Palo Alto Networks' broader efforts to build defenses against frontier models.

The company has also joined Anthropic PBC's Project Glasswing, which provides early access to Claude Mythos Preview, a model reserved for vulnerability discovery and chaining of software flaws.