ItInsightsWorld Security: AI Is Finding Bugs Faster Than Patch Teams Can Respond
By James Hartley, ItInsightsWorld Security desk
A New York Times report frames a security problem that is now visible in patch queues, open-source inboxes and incident call bridges: AI-assisted vulnerability discovery is moving faster than repair work. The pressure is not limited to large technology vendors. It also reaches projects whose code is embedded in services, browsers, enterprise applications and the wider internet supply chain.
For IT managers, developers and security practitioners, the practical question is simple: what does a team do when finding flaws is cheaper, faster and more continuous than shipping trusted fixes?
October 21 and the Human Cost of an AI Discovery Surge
On October 21, Fouad Matin, OpenAI’s leader for cybersecurity readiness and security, had planned to spend the evening with friends for his birthday. He slept about two hours that day and spent the rest coordinating fixes for two severe security flaws discovered by AI. If abused, the flaws could disrupt broad internet services. The work involved security researchers at startups and major technology companies, as well as government officials from the United States and other countries.
The open list still held more AI-found vulnerabilities waiting for repair.
Matin said OpenAI has created multiple teams focused on cybersecurity and has committed substantial compute to locate and address issues. In his account, AI companies have a responsibility to deal with the risks the technology reveals; the stated goal is to scan and patch software as broadly as possible.
What the Vendor Numbers Show
OpenAI said it started using its own technology to scan open-source software in May this year. Since then, internal and external researchers have used its models to uncover more than 1,300 vulnerabilities. Open-source software can be used for free and is widely deployed across network services, supporting much of internet functionality.
Google said that in June this year it used AI to identify and fix 1,072 vulnerabilities in the Chrome browser, exceeding the total repaired in the previous two years. It also said that since the beginning of this year, the count of severe Chrome vulnerabilities fixed has surpassed the previous eight years combined.
Microsoft released 1,461 patches in October this year, above the 1,247 patches it issued during the full year of 2025.
| Company or project | Reported volume from the New York Times account |
|---|---|
| OpenAI | AI scanning of open-source software began in May this year; researchers using its models have found more than 1,300 vulnerabilities. |
| Google Chrome | 1,072 AI-discovered browser vulnerabilities were fixed in June this year, exceeding the previous two-year total. |
| Microsoft | 1,461 patches were released in October this year, above the 1,247 patches issued during the full year of 2025. |
| Node.js | 50 security vulnerabilities were confirmed and repaired this year; February report volume reached 4.6 times the normal level. |
Maintainer Experience: More Reports, Fewer Off Hours
Some maintainers believe this workload may only be starting. Matteo Collina, CTO at Platformatic and a Node.js maintainer, noted that Uber and Netflix use the tooling. “What is appearing now is only the first bit; the situation will become ten times worse,” Collina said.
The OpenJS Foundation, which supports Node.js, said the project confirmed and repaired 50 security vulnerabilities this year, more than in any previous year. In February this year, it received vulnerability reports 4.6 times the normal amount. This month, OpenAI contacted Collina about one newly discovered flaw; a potential attacker could exploit it to compromise internet systems and steal data.
Collina’s weekly work pattern is now seven days, and at night he lets AI agents continue tasks. “No one knows what this profession will become,” he said.
Why AI Changes the Economics of Vulnerability Research
Before AI-assisted scanning, finding software flaws depended heavily on human expertise. Companies employed security researchers to review code and used bug bounty programs to pay white-hat hackers for findings. AI changes the speed and value of that work. It can search continuously, keep probing where a human may stop, inspect small defects in different parts of code, and combine them into dangerous attack paths.
Volker Dusch, a security engineer with the PHP Foundation, said that two years ago finding a security vulnerability was considered a remarkable event; now the peer recognition attached to that work is fading, and the industry’s way of assessing it is changing. The PHP Foundation, involved in maintaining the PHP language, disclosed one severe vulnerability discovered by AI last week.
Flooded Reports, Low-Quality Noise and Offensive Capability
Since late last year, open-source maintainers have observed a noticeable rise in reports. Some identify serious threats, but others are AI-generated low-quality submissions that are confusing and untested. In late January this year, curl announced it stopped paying bug bounties because it could not process the volume of incoming reports.
Defenders are also concerned about AI attack potential. In April this year, Anthropic said its Mythos model had strong hacking ability and could not be safely opened to the public. It later provided the model to a group of trusted companies and government agencies through Project Glasswing, for use in inspecting and patching systems. OpenAI and Google have also given defenders AI models focused on cybersecurity.
Access is still expanding. Vals AI recently created a simulated attack task in which one model attempted to breach a fictional company. The test found that one advanced OpenAI model available only to approved defenders completed 81% of the objectives, followed by one freely accessible DeepSeek model from China that completed 73%. As such capabilities spread, security teams in Silicon Valley are accelerating internal checks, fearing attackers may discover the same flaw soon.
The PHP Case: Weeks of Coordination Behind a Single Fix
The PHP repair shows where the pressure lands. In July this year, researchers at cybersecurity firm Calif began running AI agents against PHP, which reports say is used by about 70% of websites. They found one flaw hidden in code for more than ten years that could allow attackers to run malicious code on affected sites.
Calif contacted Dusch and researchers at Wiz, a security company owned by Google. OpenAI engineers and PHP Foundation engineers also joined. The group worked for weeks in a private Slack channel and used daily calls to coordinate the fix. Calif CEO Thai Duong said the patch touched about 200 files and was an exceptionally large update.
While building the patch, they notified government agencies, Cloudflare and other organizations to increase remediation coverage. The interval between flaw discovery and widespread repair was the most anxious phase. Cloudflare CEO Matthew Prince said that during the past two months he had been in extreme fear that others would find the flaw first or that information would leak before the patch was ready. He predicted that, in the next two years, an average of one internet-wide severe vulnerability could appear every week.
The Industry Does Not Agree on What Comes Next
Security professionals are split over how to read the present moment. Jason Clinton, Anthropic’s deputy chief information security officer, wrote on October 22 that although the past four days had already been very bad, conditions would worsen again before they could improve.
Prince took a different view. He argued that finding these long-existing problems can ultimately make the internet safer. He expects the next two years to feel difficult, but says that after this repair wave, people may reach a much more secure internet.
Security Checklist for Patch and Application Teams
Based on the scenario described in the reporting, teams should stress-test their workflows for a higher volume of AI-assisted findings. The goal is not to chase every report, but to protect the path from discovery to trusted patch:
- Triage by exploitability, not novelty. Prioritize issues that can run malicious code, steal data, or affect widely deployed services such as browser, open-source infrastructure and PHP-based sites.
- Require proof for incoming AI reports. Treat confusing or untested submissions as low trust until reproduction, affected version range and impact scope are verified.
- Map internal dependencies. Identify which applications, libraries and platform services rely on open-source components that external scanners may already be examining.
- Prepare a coordinated disclosure lane. For high-severity flaws, define who talks to maintainers, governments, CDN and infrastructure partners before public details circulate.
- Track repair velocity. Compare your current patch cadence against the vendor examples above, and escalate when report intake exceeds validation capacity.
- Use AI agents carefully. If teams let AI assist overnight work, maintain human approval gates for code changes, releases and security advisories.
Common Misconceptions
| Misconception | What the reporting suggests |
|---|---|
| Every AI-generated vulnerability report is useful. | Some are valid serious threats, but many are confusing, untested or low quality, which adds burden to maintainers. |
| Restricting advanced AI models to defenders keeps findings private. | Models that can find and exploit flaws are increasingly accessible, and defenders are worried attackers may locate the same issue soon. |
| Faster discovery automatically improves security. | Repair capacity, coordination and disclosure pressure are now the limiting factors; open-source maintainers are already working around the clock. |
FAQ
Why does AI discovery create a different problem than past bug-finding? The article describes AI that can search continuously, find small defects in different code areas and combine them into attack paths. That can increase report volume faster than human validation and patching capacity.
Who bears the most immediate burden? Open-source maintainers and vendor security teams. Examples in the report include Node.js maintainers, PHP Foundation engineers, and staff at OpenAI, Google and Microsoft.
What does the PHP repair reveal about coordination? A single high-severity flaw required weeks of private collaboration among Calif, Dusch, Wiz, OpenAI and PHP Foundation engineers, plus notification to government agencies, Cloudflare and other organizations.
Is the outlook only negative? Not uniformly. Jason Clinton says the situation may worsen before improving, while Matthew Prince expects the internet to become safer after this round of fixes.
Action Advice: What to Do in the Next 30 Days
- Create a validation queue for externally reported vulnerabilities, with separate lanes for AI-assisted reports and human-confirmed exploits.
- Inventory critical open-source dependencies, especially components used by public-facing services and developer platforms.
- Set clear severity rules: prioritize flaws that enable code execution, data theft or broad service disruption, as discussed in the Node.js, PHP and Chrome examples.
- Coordinate with vendors and infrastructure partners early when a serious flaw affects shared services, while limiting public disclosure until patches are ready.
- Review staffing and automation limits so AI-assisted triage supports humans instead of replacing release approval.
Prepared for ItInsightsWorld Security. Contact: contact@itinsights.com.