an AI just found 14,090 real security bugs in open source — in two months

Unit 42, Palo Alto Networks' threat research arm, spent two months pointing an ensemble of frontier AI models at almost four thousand open source software projects. It didn't ask the models to summarize code or write documentation. It asked them to find real, exploitable vulnerabilities — the way a human security researcher would, just at a scale no human team could match. The result: 14,090 confirmed vulnerabilities, and 99.4% of them had never been reported before.

How NOVA actually works

NOVA runs without a human in the loop until the final review stage. It reads source code, reasons about what a function is actually meant to do versus what it actually does, builds a working proof-of-concept exploit to confirm the bug is real rather than theoretical, and generates a disclosure report — all autonomously. Unit 42's own report says no single model found everything; each frontier model in the ensemble surfaced a distinct set of bugs the others missed, which is itself a data point about how differently these models currently reason about code.

What kind of bugs an AI finds that a scanner doesn't

Traditional static analysis tools are good at memory bugs — buffer overflows, use-after-free, the kind of flaw with a recognizable shape in the code. NOVA's numbers flip that expectation almost entirely: just 4% of what it found were memory-safety issues. The other 92% were semantic and logic flaws — an access control check that's present but wrong, a path traversal filter that misses one encoding, an SSRF hole that only opens under a specific request sequence. The project-by-project language breakdown backs this up: JavaScript/TypeScript projects (2,197 of them) yielded 2,836 findings; Go projects (1,636) yielded 3,281; and a much smaller set of just 17 PHP projects still produced 2,740 findings — a sign of how concentrated logic-flaw density can be once you have a reasoning system patient enough to actually trace the logic end to end.

The number that should actually worry you: the patch window

The patch window has collapsed.

Unit 42, Frontier AI Vulnerability Burst report

Security teams have spent two decades planning around a rough rule of thumb: attackers typically take weeks to reverse-engineer a disclosed patch into a working exploit, so a 30-to-55-day patch cycle was survivable. NOVA's own existence is Unit 42's argument that this assumption no longer holds — if an AI system can go from raw source code to a validated, working proof-of-concept in the course of an automated scan, the gap between 'a bug exists' and 'a bug is weaponized' isn't measured in weeks anymore for anyone with comparable tooling. It's the exploitation side of the same trend CISA flagged when an AI system was actively hunting a Tomcat vulnerability before the agency's own patch deadline had even passed. Unit 42's proposed fix isn't faster patching — it's 'Advanced Virtual Patching,' shielding a known-vulnerable service at the network layer within hours instead of waiting on a vendor's release cycle.

Here's a prediction, and I'll label it as one: expect a genuine spike in disclosed CVEs for boring, long-stable open source infrastructure over the next six to twelve months, as defensive research groups and attackers point similar tooling at the same haystack. It's also why the debate around the White House's AI safety evaluation framework — which the administration classified rather than published — matters more than it might seem: frontier-model vulnerability research is exactly the kind of dual-use capability that framework was reportedly built to assess. None of this is a reason to panic about any specific piece of software. It's a reason to actually run the dependency and supply-chain hygiene basics instead of treating them as optional.