PAN-OS, the software behind Palo Alto Networks’ firewalls, is getting a major update. PAN-OS 12.2 Ceres focuses on proactively protecting software through Advanced Virtual Patching, including automatically detecting vulnerabilities with the company’s proprietary autonomous AI system, NOVA.
Palo Alto Networks is aware of the so-called “cryptographic reset,” the phenomenon in which patching and vulnerability scans are needed faster than ever to stay ahead of AI-driven attacks. For now, initiatives such as Daybreak and Glasswing, from OpenAI and Anthropic, respectively, are intended to give defenders earlier access to new large language models (LLMs). However, the recent hack at HuggingFace carried out by such an AI model from OpenAI demonstrated that by no means does every organization have access to these systems and is therefore adequately protected. For Palo Alto Networks and other security providers, it is therefore crucial to provide more protection than ever before.
NOVA
The Palo Alto Networks research team therefore built a system it calls the Network and Open-Source Vulnerability Analyzer, or NOVA for short. It is an agent-based pipeline that scans code without human intervention, detects vulnerabilities, and then delivers a working proof of concept for successful exploits. A human is only involved during the final check.
The numbers are certainly impressive. Behind the scenes, NOVA has already analyzed 3,915 open-source projects across six software ecosystems and identified 14,090 confirmed vulnerabilities. Of these findings, 99.4 percent had not yet been documented anywhere. Under CVSS 4.0, nearly 40 percent were labeled “High” or “Critical.” By comparison: OSS-Fuzz, which has been running since 2016, claims to have helped identify more than 10,000 vulnerabilities as of August 2023. NOVA achieved something comparable in just two months.
That does not necessarily mean that every vulnerability was immediately exploitable. Even CVSS 4.0 scores are not entirely focused on quantifying the actual danger, but are merely an indication of the impact a specific vulnerability has. If that impact is enormous but the system is not mission-critical, a vulnerability with a much lower CVSS score can still be far more damaging.
No longer just crash bugs
What stands out is the type of flaws NOVA discovered. It goes beyond mere automated scanning. Systematic searches for vulnerabilities without human intervention are well known. Fuzzers have been hunting for memory corruption and integer overflows for years. However, that category accounted for only 4 percent of NOVA’s findings. The remaining 92 percent consisted of logical and semantic errors, such as flawed access control, path traversal, code injection, and server-side request forgery.
Each ecosystem was found to have its own weak spot. C and C++ often face memory safety issues, while JavaScript and TypeScript are prone to code injection and prototype pollution; PHP, Java, and Python, on the other hand, primarily struggle with authorization. The supply chain also came into focus: 1,280 vulnerabilities in dependency packages led to over 4,000 downstream exposures in applications that use such packages.
The patch window has collapsed
Palo Alto concludes from this that “time-to-patch” has collapsed. Whoever finds vulnerabilities faster automatically reduces the time between discovery and exploitation. This was previously complicated by “security by obscurity”: if you are the only one aware of a vulnerability, it is, as far as anyone knows, not immediately dangerous. Cybercriminals and state actors can now exploit and discover cyber threats much more quickly. Attackers don’t even need the latest Frontier model to reverse-engineer patches. Incidentally, this also applies to defenders: during the recent OpenAI incident, HuggingFace attempted to use Claude Fable or GPT-5.6 Sol for forensic assistance, but only Z.ai’s open-weight model, GLM-5.2, aided the investigation because it lacked the guardrails employed by OpenAI and Anthropic.
Nevertheless, those American AI labs are also trying to support the defenders wherever possible. According to CERT-EU, models such as OpenAI’s Codex Security and Anthropic’s Claude have shifted the role of AI from an assistant to human auditors to the primary discoverer of zero-days. Anthropic reported in early 2026 that it had found more than 500 vulnerabilities in open-source software using Opus 4.6, which is already considered obsolete after only a short time. The Cloud Security Alliance also warns that the existing CVE infrastructure is buckling under this flood of reports, with a forecast of about 66,000 CVEs this year.
Advanced virtual patching as the solution
Palo Alto’s solution lies in Advanced Virtual Patching, a feature of the new PAN-OS 12.2 Ceres release, which the company says includes over 55 innovations. The idea is to provide protection within a few hours, even before an official patch is available. With this, the company aims to reduce the average “exposure window” of 55 days to nearly zero.
Virtual patching itself is a familiar term, albeit now with a new twist. Palo Alto previously recommended temporarily mitigating PAN-OS vulnerabilities using threat prevention signatures and by shielding management interfaces, while awaiting an actual update. This approach is primarily intended for environments that cannot simply be taken offline, such as OT, critical infrastructure, and healthcare equipment. A new detection engine called “vaulted protection” is designed to roll out that protection securely.
Then there’s another issue: AI costs. Although LLMs are capable of helping, the price tag for security is something that gives wealthy attackers an advantage over defenders with limited budgets. Palo Alto CEO Nikesh Arora warned just last month that the cost of AI tokens must come down significantly for widespread adoption, the company itself says it spends about a million dollars a day on tokens. The fact that NOVA relies on an “ensemble” of multiple frontier models simultaneously suggests that the costs are substantial. There’s a good reason for that, though.
A test conducted on 14 projects revealed that different AI models consistently found different vulnerabilities in the same code. The model with the highest yield found 235 confirmed vulnerabilities, 185 of which were not detected by any other model. For responsible disclosure, Palo Alto collaborates with open-source maintainers and so-called clearinghouses, initiatives that catalog and eliminate exploitations of vulnerabilities at scale. The company says it continues to invest in AI-native research and collaboration with partners.