Unit 42 implements GPT-5.6 Daybreak for its clients and identifies 36% of exposures outside the CVE catalog

Palo Alto Networks has been deploying specialized artificial intelligence models for cybersecurity within its clients’ systems since August 12. The company’s Unit 42 team utilizes these models to identify potential attack vectors and assess their exploitability. Notably, 36% of the exposures identified do not correlate with any known vulnerabilities.

Palo Alto Networks has expanded its Unit 42 Frontier AI Exposure Analysis service, integrating OpenAI’s most advanced cybersecurity model, GPT-5.6 Daybreak, directly into client environments. Previously, this model was not available for commercial use. The announcement was made by Sam Rubin, who leads advisory and threat intelligence operations at Unit 42.

This expansion builds on a service launched in May 2026 under the Frontier AI Defense banner. According to Palo Alto Networks, it has already presented its findings to over a thousand security teams and deployed the service with several hundred clients. Notably, just two days prior, OpenAI had initiated a program allowing authorized partners to utilize its cybersecurity models, with Palo Alto Networks being one of the participants.

Model Identifies Vulnerabilities and Validates Exploitability

Unit 42 has traditionally conducted penetration testing through manual exploration by specialists, who seek weaknesses and attempt to chain them together to access data or privileges. The newly described service leverages models to execute part of this assessment. The AI traverses applications, identities, infrastructure, and cloud configurations, testing the pathways it identifies.

This process is divided into five key functions, according to the company. The first applies the latest cyber models to search for exposures. The second allocates tasks across multiple models, each optimized for specific tasks. The last three functions address vulnerability discovery, attack simulation, and remediation planning. Unit 42 experts validate the findings before conclusions are drawn. This human validation is presented as a critical aspect of the service, as the model generates hypotheses that can only be confirmed by an analyst.

One Third of Identified Exposures Are Not Listed in Vulnerability Catalogs

Security teams typically rely on the Common Vulnerabilities and Exposures (CVE) catalog, which documents known software vulnerabilities. Scanners compare installed versions against this list to flag missing patches. Palo Alto Networks asserts that 36% of the exposures identified through its service do not correlate with any entries in the CVE catalog.

The company attributes this finding to the nature of the identified vulnerabilities, which often involve several minor flaws that, when combined, can yield results that none of them could achieve individually. A scanner that checks versions may overlook these since it lacks the capability to analyze them in combination. Palo Alto Networks does not disclose the specific scope or methodology of this internal measurement; however, it aligns with long-standing observations from penetration testing teams that combining minor flaws can expose more pathways than isolated critical vulnerabilities.

Clients Receive Results Without Direct Access to the Model

OpenAI maintains strict controls over how authorized partners use its cybersecurity models. The approach ensures that model access remains with the authorized partner and is never transferred to the end client. Therefore, a security department subscribing to this service receives findings and a remediation plan without direct interaction with the model itself.

This structure places the provider between the enterprise and the model supplier. It also extends to tool selection, as Palo Alto Networks notes that no single model is suitable for all security tasks, meaning multiple models operate behind the same framework. Clients, therefore, cannot discern which model produced which findings. This complexity is compounded by uncertainty about the versions in use; OpenAI’s partner program pages reference only GPT-5.5, whereas Palo Alto Networks introduces GPT-5.6 Daybreak.

Raphaël Marichez, the security director for France and Southern Europe at Palo Alto Networks, positions this offering as a counterbalance to the disparity between attackers and defenders. This argument calls for careful consideration, given that the capabilities described can effectively be employed both to identify vulnerabilities and to exploit them—a capability that offensive AI mapping had already established. Security leaders contemplating this service should ensure their contracts clearly stipulate three critical areas: the exact scope of testing, retention of findings, and the handling of data processed by the model.

Scroll to Top