Silicon Valley is grappling with a flood of bugs found by AI

SAN FRANCISCO >> On Oct. 21, Fouad Matin, who leads OpenAI’s cyber-preparedness and safety efforts, had planned to spend the evening with friends to celebrate his birthday.

Instead, he was running on about two hours of sleep as he juggled efforts to patch two critical security vulnerabilities that had been discovered by artificial intelligence, which threatened to upend much of the internet. The work had entangled security researchers from startups and tech giants, as well as U.S. and international government officials.

More bugs that had been found by AI were also in Matin’s queue, and they needed fixes. OpenAI has several teams devoted to cybersecurity, and it is pouring significant computing power into finding and fixing problems, Matin said.

AI “really opens this Pandora’s box of vulnerabilities,” he said. AI companies have a responsibility to track down and fix problems, he added, as “the true goal is to scan and patch as much of the world’s software as possible.”

Across Silicon Valley, technologists like Matin are grappling with a major shift in cybersecurity. AI has proved so skillful in uncovering software flaws that engineers are dealing with a sudden flood of problems that they are struggling to fix before they can be exploited and abused by hackers. Companies are throwing more time and resources at handling the vulnerabilities, and some technologists are panicking.

The volume of flaws has been overwhelming. In May, OpenAI began using its technology to scan for flaws in open-source software, which is widely used because it is free and underpins much of the internet. So far, the company said its models had been used by internal and external researchers to find more than 1,300 bugs.

In June, Google also used AI to find and fix 1,072 bugs in its web browser, Chrome, which was more than it had patched in the last two years. So far this year, it has fixed more severe bugs than the last eight years combined. And last month alone, Microsoft published 1,461 patches, more than the 1,247 it issued in all of 2025.

It is likely only the start.

“This is the first trickle that is coming through, and it is going to get 10 times worse,” said Matteo Collina, chief technology officer of a startup called Platformatic, who also helps maintain Node.js, a free developer tool that is used by tech companies like Uber and Netflix.

Node.js has been bombarded with bug reports and has confirmed and fixed 50 vulnerabilities this year, more than any other year in its history, according to the OpenJS Foundation, the nonprofit that hosts Node.js. Last month, OpenAI contacted Collina to warn him that a flaw it had discovered could let potential attackers break into systems across the internet and steal data.

To deal with the surge, Collina said he has been working seven days a week and has AI agents working on his behalf overnight. “No one knows what this profession will look like anymore,” he said.

(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to AI systems. The two companies have denied those claims.)

Bug hunting was once a human art. Tech giants hired security researchers to comb through their code, and they created programs called “bug bounties” that paid white-hat hackers to disclose vulnerabilities.

“Two years ago, it was really cool to find a security vulnerability,” said Volker Dusch, a security engineer who works with the PHP Foundation, a collective that maintains the programming language PHP. The group disclosed last week that AI discovered a critical vulnerability in its code. “That social credit is kind of gone. The whole way this is valued is shifting.”

AI has rapidly outstripped human hackers. The technology can hunt relentlessly for flaws, continuing to work when humans would likely give up. And it can search broadly for vulnerabilities, chaining together lots of little flaws in far-flung bits of code to build a dangerous attack.

The technology’s effect on security became evident late last year. Makers of open-source software noticed a steep rise in bug reports. Some reports were slop — messy, untested reports of security flaws that were AI-generated — but others described significant threats.

By late January, a small open-source project known as Curl said it would stop paying bug bounties, citing the deluge of reports. By February, Node.js received 4.6 times more reports than usual.

In April, fears of an AI-incited internet meltdown heightened. Anthropic said it had created an AI model called Mythos that was so competent at hacking that it could not safely be released to the public. Instead, Anthropic granted access to a trusted group of companies and governments, called Project Glasswing, so they could begin scanning their systems for holes and patching them. OpenAI and Google also developed AI models focused on cybersecurity and shared them with defenders.

But AI models that can find and exploit vulnerabilities have become widely accessible. Vals AI, a company that evaluates the performance of AI technologies, recently presented models with a cybersecurity challenge to hack into a fictional company. One of OpenAI’s most advanced models, which can be used only by approved defenders, completed 81% of the objectives in the test, Vals AI said. It was followed by a freely available model from DeepSeek, a Chinese startup, which completed 73% of the assigned tasks.

The availability of powerful AI has fueled a frenzy in Silicon Valley to find and fix vulnerabilities before attackers get to them. That has turned many tech jobs from building new technologies to defending existing ones from collapse.

In July, researchers at a cybersecurity firm called Calif turned their attention — and their AI agents — to PHP, which is used to build about 70% of websites. They uncovered a flaw, hidden in the code for more than a decade, that could allow malicious code to run on any of those sites.

The researchers connected with Dusch and with peers at Wiz, a security company owned by Google, who had uncovered a similar vulnerability. Together, they began working to fix it, along with others from OpenAI and the PHP Foundation. Engineers huddled for weeks in a secret Slack channel and held daily phone calls to coordinate a repair for the problem.

Fixing the flaw required changes to about 200 files, said Thai Duong, CEO of Calif, which he called “a massive patch.” As they prepared the fix, the group notified organizations that could help deploy it widely, including governments and Cloudflare, an internet infrastructure company.

“The last two months, I’ve been living in a state of abject terror that someone else would find the bug or the information would leak before everyone had patches and fixes in place,” said Matthew Prince, CEO of Cloudflare. “For the next two years, we’re going to be living in a state where, on average, every week there is a critical, internet-wide vulnerability.”

For some security experts, the recent spate of bugs is a sign of impending doom.

“As bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better,” Jason Clinton, the deputy chief information security officer at Anthropic, wrote in a social media post Oct. 22.

Others see light at the end of the tunnel.

“You’re actually safer in a world where we are finding these things,” Prince said. “The next two years, it’s going to feel like the sky is falling, but I think we come out the other end of it with a much safer internet.”

© 2026 The New York Times Company