Abstract:
According to the New York Times, AI is discovering software security vulnerabilities at an unprecedented rate, and engineers responsible for fixing them are finding it increasingly difficult to keep up. Tech companies and open-source software maintainers are rushing to patch the vulnerabilities before hackers can exploit them to protect the software that powers a wide range of websites and online services.
On October 21, Fouad Matin, who is responsible for OpenAI network security preparation and security work, originally planned to celebrate his birthday with friends in the evening.
But that day, he only slept for about two hours and was busy coordinating the repair of two serious security vulnerabilities discovered by AI. Once these vulnerabilities are exploited, they may have an impact on a wide range of Internet services. Those involved in the processing include security researchers at startups and tech giants, as well as government officials in the United States and other countries.
On his to-do list, there are more loopholes discovered by AI waiting to be fixed.
Mateen said that OpenAI has arranged multiple teams to be responsible for network security and invested a lot of computing power to find and fix problems. He believes that AI companies have a responsibility to address the risks revealed by this technology, with the ultimate goal of scanning and patching the world's software as much as possible.
The speed of discovering vulnerabilities has increased significantly, which is putting unprecedented pressure on those who fix the vulnerabilities.
According to OpenAI, the company began using its own technology to scan open source software in May this year. Since then, internal and external researchers have discovered more than 1,300 vulnerabilities using its model. This type of software can be used for free, is widely deployed in various network services, and supports a large number of functions of the Internet.
Google also used AI to discover and fix 1,072 vulnerabilities in the Chrome browser in June this year, exceeding the total number fixed in the previous two years. Since the beginning of this year, Chrome has fixed more serious vulnerabilities than in the previous eight years combined.
Microsoft released 1,461 patches in October this year alone, exceeding the 1,247 patches for the whole of 2025.
In the eyes of some software maintainers, this intensity of work may have just begun.
Matteo Collina is the chief technology officer of the startup Platformmatic and is involved in maintaining the open source development tool Node.js. Tech companies such as Uber and Netflix are using this tool.
"What is happening now is just the first bit, and the situation will be ten times worse." Colina said.
According to the OpenJS Foundation, a non-profit organization that supports Node.js, Node.js has identified and fixed 50 security vulnerabilities this year, more than any year in its history. In February this year, the project received 4.6 times the usual number of vulnerability reports.
This month, OpenAI contacted Colina again to alert him to a newly discovered vulnerability. Potential attackers may use it to invade systems on the Internet and steal data.
In order to handle the increasing number of reports, Colina now works seven days a week and lets the AI agent continue to handle tasks for him at night.
"No one knows what this profession will become in the future." He said.
In the past, finding software vulnerabilities was a job that relied heavily on personal experience. Tech companies hire security researchers to review code and also pay white hat hackers to find and report problems through "bug bounty" programs.
AI is changing the speed and value of this work.
It can keep searching, picking up where human researchers might have given up. It can also examine small flaws scattered across different codes and combine them into dangerous attacks.
Volker Dusch, a security engineer at the PHP Foundation, feels this clearly. The foundation, which maintains the programming language PHP, last week disclosed a serious vulnerability discovered by AI.
Dusch said that two years ago, discovering a security vulnerability was a big deal. Now the peer recognition it brings is waning, and the way the entire industry evaluates this work is changing.
However, the increase in vulnerability reporting also comes with a significant additional burden.
Since late last year, open source software maintainers have noticed a significant increase in the number of reports. Some of them do reveal serious threats, others are low-quality AI-generated reports that are confusing and untested.
In late January this year, the open source project curl announced that it would stop paying bug bounties because it was unable to cope with the influx of reports.
At the same time, the AI's offensive capabilities also make the defenders increasingly nervous.
In April this year, Anthropic stated that the Mythos model it developed had strong hacking capabilities and could not be safely opened to the public. The company then made the model available to a group of trusted businesses and government agencies to inspect and patch systems through a program called Project Glasswing. OpenAI and Google also provide cybersecurity-focused AI models to defenders.
But models that can find and exploit vulnerabilities are becoming increasingly available.
The AI evaluation company Vals AI recently set up a simulated attack task, requiring the model to invade a fictional company. The test results showed that an advanced model from OpenAI, which is only open to approved defenders, completed 81% of the test goals; followed closely by a freely available model from the Chinese company DeepSeek, with a completion rate of 73%.
As such capabilities proliferate, security teams in Silicon Valley are speeding up investigations. They worry that the same vulnerability may soon be discovered by attackers.
A PHP repair process shows who exactly this pressure falls on.
In July of this year, researchers at cybersecurity company Calif began letting AI agents inspect PHP. According to reports, about 70% of websites use this programming language. A research team discovered that a vulnerability that had been hidden in the code for more than a decade could allow attackers to run malicious code on affected websites.
Calif then contacted Dusch and researchers at Wiz, a Google-owned security company that discovered similar vulnerabilities. Engineers from OpenAI and the PHP Foundation are also involved.
The team collaborated in a confidential Slack channel for weeks, coordinating fixes over the phone every day. Calif CEO Thai Duong said the fix involves about 200 files and is a massive patch update.
While preparing the patch, they also notified government agencies, Internet infrastructure company Cloudflare and other organizations, hoping to expand the repair coverage as much as possible.
The period when the vulnerability has been found but has not been extensively repaired has become the most stressful stage for participants.
Cloudflare CEO Matthew Prince said he has spent the past two months in a state of extreme fear that someone else would discover the vulnerability early, or that information would be leaked before a patch was ready.
He predicts that in the next two years, on average, a serious vulnerability will appear every week that affects the entire Internet.
In the security industry, there are different judgments on the next situation.
Anthropic deputy chief information security officer Jason Clinton said on social media on October 22 that although the situation in the past four days has been bad, it will get worse in the future before it may get better.
Prince believes that uncovering these long-standing problems will ultimately make the Internet safer. He predicts that the next two years will be very bad, but after this round of repairs is completed, people can expect a much safer Internet.
Comments