AI Is Exposing More Software Flaws Than Security Teams Can Fix

Artificial intelligence is already driving a sharp rise in discovered software vulnerabilities, and the pace of that vulnerability surge is now outstripping the ability of many companies and open source maintainers to patch them. The risk is not a hypothetical future scenario: it is happening now through widely available AI tools, including open-weight models, which are helping researchers and attackers find flaws faster than ever.

That is the central warning emerging from the latest security data. While much of the public conversation around AI has focused on extreme long-term dangers, the cybersecurity effect that matters most today is more immediate: more vulnerabilities are being found, disclosed and counted, and human defenders are struggling to keep up.

Recent figures from major software vendors illustrate how intense the pressure has become. Microsoft said it had already issued 974 patches for common vulnerabilities and exposures, or CVEs, in the current month, a record pace. Oracle reported 1,448 patches in July, far above the 309 it shipped in the same month a year earlier. Google Chrome’s two major releases in June carried 1,072 fixes, exceeding the total number of vulnerability patches the browser had shipped across the previous 23 major releases combined.

At the same time, outside researchers are using AI to uncover flaws in large numbers. Mozilla said in April that a bug-hunting sprint using Anthropic’s Mythos model turned up 271 vulnerabilities in Firefox. Those disclosures do not necessarily mean software is becoming instantly less secure in every case, but they do show that the vulnerability detection pipeline is accelerating in a way security teams cannot ignore.

Why the AI vulnerability surge matters now

The immediate significance of the trend is simple: the number of known flaws is growing faster than the people responsible for fixing them can often respond. That creates a widening gap between discovery and remediation, which can leave organizations exposed even when the problem has been identified.

Security researchers have long known that software systems contain huge numbers of flaws. What is changing is the rate at which those flaws are being found. AI tools are lowering the cost of scanning code, testing assumptions and exploring attack paths, which means more bugs can be surfaced in less time.

That can help defenders, but it also creates pressure. Every newly discovered vulnerability has to be triaged, verified, prioritized and patched. In enterprise environments, patches must then be tested and deployed without breaking critical systems. In open source projects, a small group of volunteers may be responsible for maintaining software used by thousands of downstream products.

The National Cyber Security Centre in Britain has warned that simply identifying flaws does not equal protection. The real challenge begins after discovery, when teams have to turn findings into actual fixes.

“Just finding vulnerabilities does nothing to improve your security,” the UK National Cyber Security Centre has said, underscoring the gap between discovery and defense.

How big is the increase in vulnerabilities?

The scale is striking. According to Jerry Gamblin, head of research at Empirical Security and founder of RogoLabs, the cve.icu project had recorded 66,401 CVEs by Wednesday this week. That is roughly double the 33,512 CVEs logged by the same point last year, showing how quickly the known vulnerability universe has expanded.

The comparison with earlier years is even more dramatic. In all of 2022, the year OpenAI released the first version of ChatGPT, cve.icu recorded about 25,000 CVEs. The current pace suggests a much larger and faster-moving security landscape than the one many teams were built to handle.

It is important, however, not to confuse more CVEs with more harm in a direct one-to-one sense. A CVE is a publicly identified and confirmed software flaw, not a measure of damage by itself. More listings can also reflect better detection and disclosure.

Gamblin argues that a larger CVE count does not automatically mean a larger security crisis, noting that more recorded flaws can also indicate the system is working as intended by surfacing known issues.

Even so, the practical burden is obvious. Each additional disclosure adds work for vendors, integrators, administrators and security analysts. If those teams cannot move at the same speed as discovery, attackers may gain time and opportunity.

What is driving the surge?

The clearest driver is AI-assisted vulnerability research. Mainstream models and open-weight systems can accelerate code review, pattern recognition and exploit discovery, especially when paired with human expertise. That makes bug hunting faster, cheaper and potentially more scalable than traditional manual methods alone.

Researchers have been able to use AI to identify classes of weaknesses that might otherwise have taken longer to surface. Vendors and bug bounty teams are also experimenting with AI to review code bases, prioritize alerts and generate proofs of concept. In short, the same tool that helps defenders can also help offenders.

There is also a broader structural explanation: modern software is sprawling, interconnected and often maintained under resource constraints. A single flaw can affect multiple products, libraries or services. When AI improves the speed of discovery, it does not reduce the complexity of patching, coordinating and deploying fixes.

That is why many experts see the current moment as less of a sudden AI apocalypse and more of an acceleration of a problem that already existed. The software ecosystem was vulnerable long before large language models became mainstream. AI is now increasing the velocity of discovery on top of an already strained system.

What role do open source maintainers play?

Open source maintainers are on the front line because they often support software that is widely reused but lightly staffed. A major vulnerability in an open source library can affect countless downstream products, yet the team responsible for the original code may be small and volunteer-driven.

That imbalance matters because the AI era is not just about finding more bugs. It is also about increasing the demand for attention from the very people least able to absorb it. When maintainers are overwhelmed, patches may take longer, advisories may be delayed, and downstream users may remain exposed even after a flaw is known.

Security experts increasingly describe remediation as the scarce resource. Discovery can scale with compute. Fixing software, by contrast, depends on human labor, coordination and institutional capacity.

How are security teams responding?

Security teams are responding with a mix of automation, prioritization and triage. Many organizations are using AI to help classify alerts, spot likely high-risk issues and search for patterns across code or infrastructure. The hope is that if AI helps produce more findings, it can also help sort the most urgent ones.

But there is no easy equalizer. The volume of alerts and disclosures can itself become a form of pressure. Even when a vulnerability is real, not every issue can be fixed at once. Teams must decide what is exploitable, what is exposed, what is externally reachable and what should be patched first.

That process is already difficult in large enterprises. It becomes even harder when patch cycles are slow, legacy systems are deeply embedded and security staffing is thin. For smaller companies and volunteer projects, the challenge is more acute.

Matthew Olney, Cisco Systems’ director of threat intelligence, said the industry is still figuring out where AI fits on both sides of the security contest, with attackers and defenders alike testing its value.

Olney described a landscape in which both adversaries and legitimate security teams are trying to determine where AI can give them the biggest advantage.

What do the numbers mean for companies?

The numbers suggest companies should expect more vulnerability reports, not fewer. That has direct implications for patch management, incident response, compliance and vendor risk management. A company that assumes its software exposure is static may find itself unprepared for the volume and speed of AI-era disclosures.

For enterprise leaders, the message is not that AI has made every system instantly less safe. Rather, it is that the environment has become noisier, faster and more resource-intensive to defend. Teams that already struggled to keep up with patching now face a larger intake of issues to evaluate.

That burden may extend beyond security departments. Product teams, legal teams, infrastructure engineers and executives all become part of the decision-making chain when a major vulnerability is disclosed. The result is more cross-functional pressure and a greater need for disciplined response planning.

Key pressures on defenders

• Patch overload: More disclosed flaws mean more fix queues.
• Staffing gaps: Security teams and maintainers often lack the personnel to respond quickly.
• Testing delays: Patches must be validated before release to avoid outages.
• Downstream risk: One vulnerability can affect many products built on the same component.
• Attack acceleration: Offenders may also use AI to search for exploitable weaknesses.

Could an AI slowdown solve this problem?

No. Even if governments or companies agreed to slow frontier AI development, that would not reverse the current vulnerability surge created by tools that are already available. The security impact is not waiting for some distant future model; it is being produced by existing systems today.

That distinction matters because it changes the policy debate. Much of the public discussion around AI safety has focused on speculative worst-case outcomes, including the possibility of catastrophic misuse by highly advanced systems. Those scenarios may still matter, but they do not address the immediate and measurable problem of rising vulnerability discovery.

In other words, a pause on the most advanced model training might influence future risk, but it would not remove the flood of bug findings now feeding the security pipeline. Organizations will still need to cope with the consequences of the tools already in circulation.

How does this compare with the pre-AI era?

The pre-AI security world was already dealing with a massive volume of flaws, delayed patching and uneven defensive maturity. AI has not created those weaknesses from nothing. Instead, it has intensified the tempo and raised the stakes.

Before large language models became common, security researchers could still uncover huge numbers of flaws through manual review, fuzzing, scanning and exploit development. Those methods remain important. The difference is that AI can accelerate parts of the workflow and allow a greater number of attempts, which increases throughput across the discovery pipeline.

That shift is visible in the data. The growth from 25,000 CVEs in all of 2022 to more than 66,000 by mid-September this year reflects not just more vulnerabilities being found, but also a broader industrialization of the hunt.

For defenders, the lesson is that the old baseline no longer applies. Security planning must now assume that both legitimate researchers and malicious actors can operate at AI-enhanced speed.

What happens next?

The most likely near-term future is continued escalation in both discovery and defense. AI will keep helping researchers surface new flaws, and companies will keep racing to patch them. The central uncertainty is whether remediation can keep pace.

If it cannot, the security system risks becoming increasingly reactive, with organizations waiting for disclosures rather than preventing them. If defenders can adapt, AI may still improve security outcomes by automating parts of the review and response process. The balance will differ by company, by product and by industry.

For now, the clearest conclusion is that the vulnerability explosion is no longer theoretical. It is measurable, it is accelerating, and it is happening within the limits of the current software ecosystem. The challenge for security leaders is not to debate whether the surge exists, but to decide how to absorb it.

As Gamblin put it, discovery scales with compute, while remediation scales with people — and people are much harder to scale quickly. That is the practical crisis now facing the security industry.