AI Vulnerability Discovery Goes Industrial: The Numbers Behind the Patching Crunch
Three figures, read together, describe the state of internet security more precisely than any single headline: more than 1,300 vulnerabilities located with OpenAI's own models since May, 1,072 Chrome flaws found and fixed with AI in June alone, and 1,461 Microsoft patches released in October against 1,247 for all of 2025. Set on one curve, they show a discovery capacity that has gone industrial while the human side of the chain — triage, verification, patching, distribution — has not moved at the same pace. Reporting by The New York Times maps how that gap is reshaping the daily work of security engineers at large companies and inside open source, and why several of them expect conditions to worsen before they improve.
One Night Defined by Two Flaws
On October 21, Fouad Matin, who oversees cybersecurity preparedness and safety at OpenAI, had planned an ordinary evening: a birthday celebration with friends. It turned into roughly two hours of sleep in a day spent coordinating the repair of two serious vulnerabilities uncovered by AI. If exploited, either one could have rippled across a wide span of internet services. The response pulled in security researchers from startups and from tech giants, alongside government officials in the United States and other countries. The to-do list still carried more AI-discovered flaws awaiting repair.
Matin said OpenAI has assigned multiple teams specifically to cybersecurity and is committing large amounts of compute to finding and fixing problems. In his framing, AI companies carry a responsibility to handle the risks their technology exposes, with the eventual goal of scanning and patching as much of the world's software as possible.
What the Reported Output Actually Looks Like
The speed of discovery has climbed steeply, and that is the direct source of pressure on the people who write and ship fixes. The relevant data points, as reported:
| Organization | Time frame | Reported output | Baseline given in the reporting |
|---|---|---|---|
| OpenAI | Since May this year | Internal and external researchers have used its models to find more than 1,300 vulnerabilities, with the company applying its own technology to scan open-source software | Open-source software is free to use, widely deployed across web services and underpins a large share of internet functionality |
| June this year | 1,072 vulnerabilities in the Chrome browser found and fixed using AI | More than the total repaired in the previous two years combined | |
| Since the start of this year | Critical Chrome vulnerabilities fixed | More than the total of the previous eight years combined | |
| Microsoft | October this year | 1,461 patches released | More than the 1,247 patches across the whole of 2025 |
| Node.js (OpenJS Foundation) | This year | 50 security vulnerabilities confirmed and fixed | More than in any previous year of the project's history |
| Node.js (OpenJS Foundation) | February this year | Vulnerability reports received | 4.6 times the usual volume |
The Load Lands Hardest on Small Maintainer Teams
Some software maintainers believe this intensity is only an opening phase. Matteo Collina is chief technology officer at the startup Platformatic and also helps maintain Node.js, the open-source development tool used by tech companies including Uber and Netflix.
"What we are seeing now is only the very first trickle — the situation will get ten times worse," Collina said.
The OpenJS Foundation, the non-profit behind Node.js, reported the 50 confirmed and fixed vulnerabilities this year and the February surge in reports to 4.6 times normal. This month, Collina says, OpenAI contacted him again to flag a newly discovered flaw: a potential attacker could use it to breach systems on the internet and steal data.
To keep up with the rising volume of reports, he now works seven days a week and leaves AI agents running tasks for him overnight. "Nobody knows what this job will even look like anymore," he said.
Why Machines Change the Economics of Finding Bugs
Vulnerability hunting used to depend heavily on individual experience. Companies hired security researchers to read code, and they also ran bug bounty programs that paid white-hat hackers to find and report problems.
AI is changing both the speed and the value of that work. The reported shift comes down to a few capabilities:
- It can search continuously rather than in shifts.
- It keeps trying in places where a human researcher would give up.
- It can inspect small defects scattered across different parts of a codebase and assemble them into a dangerous attack path.
Volker Dusch, a security engineer at the PHP Foundation — the body that helps maintain the PHP programming language — has felt this directly. His foundation disclosed a severe vulnerability found by AI last week. Two years ago, Dusch said, finding a security vulnerability was a significant accomplishment; the professional recognition that came with it is now fading, and the industry's way of evaluating the work itself is changing.
More reports also mean more uncompensated overhead. Since late last year, open-source maintainers have noticed a clear rise in submissions. Some genuinely reveal serious threats; others are low-quality, AI-generated reports — garbled and untested.
In late January this year, the open-source project curl announced it would stop paying bug bounties, citing its inability to cope with the flood of incoming reports.
Case in Point: One PHP Flaw, Weeks of Secret Coordination
The repair cycle for PHP shows exactly where the strain falls. In July this year, researchers at the cybersecurity company Calif began pointing AI agents at PHP, a programming language reportedly used by about 70% of websites. The team found a flaw that had been hiding in the code for more than ten years and that could let an attacker run malicious code on affected sites.
Calif then contacted Dusch, along with researchers at Wiz, the security company under Google, who had found a similar flaw. Engineers from OpenAI and the PHP Foundation joined as well. The group worked for weeks inside a confidential Slack channel, coordinating the fix over calls every day. Thai Duong, chief executive of Calif, described an enormous patch affecting about 200 files.
While the patch was being prepared, the group also notified government agencies and organizations such as Cloudflare, the internet infrastructure company, hoping to widen the reach of the fix as far as possible.
The stretch between "flaw found" and "fix widely deployed" was the most tense period for everyone involved. Cloudflare CEO Matthew Prince said he had spent the past two months in a state of extreme fear — worried that someone else would find the vulnerability first, or that details would leak before the patch was ready.
Capability Is Spreading Faster Than It Is Being Restricted
Defenders are nervous for a second reason: the same models that surface weaknesses can also exploit them. In April this year, Anthropic said its Mythos model had strong hacking capability and could not be opened to the public safely; the company instead routed it through a program called Project Glasswing, making it available to a group of trusted companies and government agencies for inspecting and patching systems. OpenAI and Google have also supplied security-focused AI models to defenders.
Even so, models able to find and exploit vulnerabilities are becoming easier to obtain. The benchmarking firm Vals AI recently set up a simulated attack task in which models had to break into a fictional company. Results reported from the test:
- An advanced OpenAI model, open only to approved defenders, completed 81% of the test objectives.
- A freely available model from the Chinese company DeepSeek followed at 73%.
As that capability diffuses, security teams across Silicon Valley are accelerating their own sweeps, worried that the same flaw will be discovered by an attacker shortly after it is found by a defender.
Two Readings of the Next Two Years
Industry judgment about what comes next is split. Jason Clinton, deputy chief information security officer at Anthropic, wrote on social media on October 22 that although the past four days had already been bad, conditions would deteriorate further before any relief.
Prince takes the longer view: surfacing these long-standing problems should ultimately make the internet safer. He expects the next two years to feel very bad, with an internet-wide severe vulnerability appearing on average every week, but anticipates a far more secure internet once this round of remediation is complete.
FAQ: Questions Readers Are Asking
What changed, technically?
Nothing about the patches themselves got easier. What changed is discovery: models can search continuously, retry where humans quit, and combine small, dispersed defects into a viable attack — which is why OpenAI reports more than 1,300 findings since May and Google reports 1,072 Chrome fixes in June alone.
Why does more reporting create risk instead of safety?
Because the gap between discovery and remediation is now an exposure window. The PHP case made that visible: a flaw hidden for more than ten years, found in July, still required weeks of covert coordination across Calif, Wiz, OpenAI, the PHP Foundation, government agencies and Cloudflare, touching about 200 files.
Who absorbs the extra workload?
Mostly maintainers of widely used open-source components. Node.js logged 50 confirmed and fixed vulnerabilities this year and a 4.6 times spike in reports in February; curl responded by suspending its bug bounties in late January; Collina now works seven days a week and runs AI agents overnight.
Are the most capable models restricted to defenders?
Partly. Anthropic withheld Mythos from public release and shared it through Project Glasswing with a group of trusted companies and government agencies, and OpenAI and Google have given defenders dedicated security models. But the Vals AI test suggests free access matters: a publicly available DeepSeek model reached 73% of objectives in a simulated breach, close behind an access-restricted OpenAI model at 81%.
What should enterprise readers do with this?
Follow the patch throughput, not the discovery claims. The reported trajectory — Microsoft shipping 1,461 patches in October versus 1,247 in all of 2025, Chrome fixing more critical flaws this year than in the previous eight years combined — indicates that exposure now depends on how quickly an organization can deploy fixes once a report arrives.
Checklist: Signals Worth Tracking
- Discovery-to-patch ratios at OpenAI (more than 1,300 findings since May), Google (1,072 in June) and Microsoft (1,461 patches in October).
- Report volume and quality at open-source projects: Node.js at 4.6 times normal in February, curl's bounty suspension in late January, the rise of untested AI-generated submissions since late last year.
- Access policy for offensive-capable models: Anthropic's Mythos and Project Glasswing, defender-only OpenAI models, and the availability gap highlighted by the Vals AI exercise.
- The two competing forecasts for the next two years, from Clinton's "worse before better" to Prince's weekly-critical-flaw projection followed by a safer internet.
Compiled by AI & Data News for AIdataNewsGlobal's General desk. The reporting summarized here comes from The New York Times and from the companies, maintainers and evaluators named above; no additional data has been introduced. Corrections and tips: editorial@aidatanews.com.