AI pushed this year's count of known software flaws past 100,000. Fixes are not keeping pace

The number of publicly catalogued software security flaws logged in 2026 passed 100,000 on Friday, according to a running tally by security data firm VulnCheck. That is a record, and more than double last year's total.

VulnCheck's Josh Bressers named the cause plainly. AI-driven bug hunting "is driving a majority of what we're seeing," he told The Register. On its face, that is good news: every flaw a friendly AI finds is one an attacker cannot find first.

But the numbers behind the milestone show a widening gap between finding problems and fixing them. For companies, that gap is where the new risk sits.

Finding flaws is cheap. Fixing them is not

The clearest case is Project Glasswing, the bug-hunting program Anthropic launched in April with about 50 partner companies, including Cloudflare and Cisco. Anthropic's most capable AI models scan software, much of it free open-source code that businesses build products on, and flag possible weaknesses.

Anthropic's public dashboard, last updated November 2, lists 37,966 candidate findings. Outside security research firms hired by Anthropic have reviewed 6,376 and confirmed 6,009 as real, a 94.2% hit rate. Anthropic has reported 5,087 to the people who maintain the affected software. It says 629 have been patched, "to our knowledge."

An independent reading is less flattering. Patrick Garrity, a VulnCheck researcher, tracks Anthropic's public ledger of findings. In a November 9 analysis he counted 217 entries marked fixed. Between early September and early November, total findings grew by about 11,800, while fixed entries rose by 15. Entries marked as duplicates jumped from 2 to about 1,670.

The two fix counts measure different things. Anthropic's covers patches it knows about anywhere; Garrity's counts ledger entries marked fixed. Either way, known fixes are under 2% of what the AI has flagged. Most findings have not yet been checked by outside reviewers.

Garrity also questions how Anthropic's 94.2% accuracy figure treats duplicates and withdrawn findings. Anthropic did not respond to Infosecurity Magazine's request for comment. Its dashboard says the rate includes bugs already reported elsewhere and flaws maintainers choose not to fix, and calls patch counts a lagging measure. VulnCheck sells vulnerability data, so it has its own commercial interest in the debate.

The bottleneck is people

Anthropic has acknowledged the problem. In October, as it launched a scanner that open-source projects can opt into and run themselves, it said it remained "bottlenecked on our human capacity to validate these findings," VulnCheck and Infosecurity Magazine reported. Anthropic describes that tool's reports as fully machine-generated, with no human review, and warns some may be wrong.

It also sent nearly 5,000 unverified reports to maintainers who asked for them in bulk. That shifts the sorting work from Anthropic to the maintainers, many of whom are unpaid volunteers.

One maintainer's experience shows the cost. Daniel Stenberg runs curl, a widely used tool for moving data over the internet that is built into countless products. In early November he said the project had received a batch of 45 reports from Anthropic. Of 34 published so far, 19 were not real vulnerabilities and six were confirmed as low severity. Nine are still under review.

Earlier research points the same way. In September, Jeff Williams, founder of security firm Contrast Security, told Dark Reading that an AI scan of a 2-million-line codebase cost about $315 in computing charges, while sorting through the results cost about $128,000.

What it means for buyers, boards and investors

Most of the 100,000 flaws will never be used in an attack. VulnCheck says about 1.4% of this year's flaws, or 1,363, have shown up in real-world attacks. Only two of the 442 flaws VulnCheck attributes to Anthropic have done so. Still, even a small share of a doubled total is a large number of exploited bugs.

Attackers are also moving faster. VulnCheck says the median time from a flaw's public disclosure to its first observed use in an attack fell from 120 days in 2025 to 80 days in the first half of 2026.

Three practical lessons follow. First, patch queues will keep swelling, so security teams cannot treat every new flaw as urgent. In June, US federal agencies were told to rank fixes by internet exposure, active attacks and how easily an attack can be automated, not by severity scores alone. Private companies can borrow that model.

Second, open-source code carries a hidden labor risk. The small teams that maintain it are now the choke point, and companies that depend on their code may need to fund or staff fixes themselves. Third, buyers of AI security tools should judge them on flaws actually fixed, not flaws counted.

For investors, Glasswing is a showcase for Anthropic's top models. The discoveries are real. Whether they make software safer depends on a human repair workforce that AI has not yet scaled. As Bressers put it: "We don't have a security problem anymore. We have a data problem now."