"In just two months, NOVA analyzed 3,915 open-source software projects and uncovered 14,090 confirmed vulnerabilities."
Palo Alto Networks' NOVA: an autonomous pipeline that does the whole discovery loop
Palo Alto Networks' vulnerability research team built an autonomous system called the Network and Open-Source Vulnerability Analyzer (NOVA). According to the report, NOVA performs a full audit loop for each project: review of project history, reading source code, identifying vulnerability candidates, creating working proof-of-concept (PoC) exploits, deterministic validation in a clean environment, generating patch candidates and producing disclosure reports. The pipeline scanned 3,915 projects across six major software ecosystems and produced findings in every ecosystem it tested.
Scale and surprise: 14,090 findings, 99.4% previously unreported, many severe
The system produced 14,090 confirmed vulnerabilities in two months; the report states 99.4% of those had not been reported publicly before NOVA’s work. Severity was not confined to low-impact bugs: by CVSS 3.1 standards 4,030 findings (28.6%) scored High or Critical, and by CVSS 4.0 standards 5,600 findings (39.7%) scored High or Critical. The team also found that only 85 of the NOVA discoveries matched entries already in the public record, and most of those public disclosures appeared two to eight weeks after NOVA’s discovery.
Built by Nubivance.
OSINTSights' secure edge-first architecture, AI content pipeline, and serverless ops are designed by Nubivance. We do this for clients too.
Talk to us →What NOVA found: diversified vulnerabilities and supply-chain reach
Contrary to the expectation that AI would mostly repeat prior fuzzing results, the report finds that memory and calculation issues were only a small fraction of NOVA’s output: 557 findings (4.0%) for memory and calculation, and 1,121 findings (8.0%) when resource-management and denial-of-service are included. Instead, 92% of the findings were semantic and logic flaws—access control and authorization, path traversal, code injection, prototype pollution and server-side request forgery (SSRF).
The distribution varied by ecosystem in predictable ways the authors map to common development patterns: C/C++ was dominated by memory-safety and resource-management flaws; JavaScript and TypeScript concentrated in code injection, prototype pollution and SSRF; Go skewed toward path traversal and SSRF; and PHP, Java and Python showed many access-control and authorization problems. The pipeline also identified 5,421 supply-chain findings—1,280 flaws in dependency packages and 4,141 downstream exposures—and validated exploitability from the downstream application with a working PoC for 2,776 of those downstream exposures.
How the harness works: multi-agent ensembles, containment, and independent validation
The NOVA architecture is agentic and ensemble-based. The harness gives agents structured workflows—understand the target architecture, identify where outside input enters, locate dangerous operations and determine connecting paths—and distributes tasks across discovery agents, proof-of-concept and verification agents, adversarial triagers and gatekeepers. The report stresses that multiple models are essential: in a controlled evaluation across 14 projects, different models often found different vulnerabilities, and every model contributed a large set of findings no other model found.
To manage risk at scale, the pipeline includes strict validation and layered containment: findings are replayed in isolated environments and pass independent validation before publication; the system runs behind containers, gVisor-style sandboxing, virtual machines, network egress controls and least-privilege permissions.
Advanced Virtual Patching, Project Lightwell, and the compressed patch window
Palo Alto Networks frames the defensive response around turning earlier discovery into faster protection. The company says its Advanced Virtual Patching capability can deliver network-level protections within hours and that protections are added to its Advanced Threat Prevention service for next-generation firewall and SASE products. The report states this approach can collapse the exposure window from an industry-average 55 days for traditional patch deployment to a near-zero window of exposure by delivering protections ahead of upstream patches. The authors also report partnerships with clearinghouses such as Lightwell and Akrites and collaborative work with Project Lightwell to combine rapid network protections with software remediation.
What this means for open-source maintainers, enterprise security teams, and threat actors
- Open-source maintainers: expect a rising volume of validated, PoC-backed vulnerability reports and coordinated disclosure from clearinghouses like Lightwell and Akrites; the report emphasizes upstream remediation and rapid reporting to maintainers.
- Enterprise security teams and patch teams: the NOVA findings press the case for virtual patching and rapid network-level protections (Advanced Threat Prevention and Advanced Virtual Patching) to protect assets where traditional patch deployment can take weeks.
- Threat actors: the report warns the patch window has collapsed and notes attackers "need not have access to the latest frontier AI model to reverse engineer patches and develop exploits automatically," compressing the timeframe defenders must act within.
The technical record in the report is stark: autonomous agents can now scale discovery across thousands of projects, validate exploits in isolated environments, and generate patch candidates. The defensive answer the authors offer is not to slow discovery but to speed protection—pair autonomous discovery with rapid virtual patching, collaboration with clearinghouses, and human judgment for prioritization. As Palo Alto Networks CEO Nikesh Arora has emphasized, the moment calls for cybersecurity providers to shift advantage back to defenders; NOVA demonstrates both why that is urgent and one approach to doing it.
