The number of Common Vulnerabilities and Exposures (CVEs) publicly disclosed in 2026 has passed 100,000, more than double last year’s total, according to VulnCheck, the security intelligence firm that tracks the count. The milestone, reached Friday, marks the first time annual CVE volume has crossed six figures and caps a year in which automated bug hunting has reshaped how software flaws are found, reported and patched.
VulnCheck’s tally stood at 100,222 as of about 4 p.m. UTC on Friday, said Josh Bressers, the firm’s director of coordinated vulnerability disclosure. Microsoft alone contributed 1,094 flaws on Tuesday, after adding 1,474 the previous month. Bressers attributed the surge squarely to artificial intelligence, saying the major CVE Numbering Authorities – the organizations authorized to assign identifiers to vulnerabilities – are all reporting at unprecedented rates. “There’s no denying the fact that vulnerability hunting with AI is driving a majority of what we’re seeing,” he said.
The raw count is not the only thing changing. The share of disclosed flaws known to be exploited in the wild has stayed roughly stable at about 1.4 percent, or 1,363 CVEs added to VulnCheck’s Known Exploited Vulnerabilities catalog this year. But as Bressers noted, a constant percentage of a vastly larger base means a far larger absolute number of actively attacked bugs. The median lag between publication of a CVE and observed exploitation has also compressed, falling from 120 days in 2025 to 80 days in the first half of 2026 and shrinking further since, with some attacks following disclosure by only hours.
CrowdStrike found that 88 percent of the exploitation it observed from January through June using public proof-of-concept code occurred within 48 hours of that code’s release. Groups linked to China, including Vault Panda and Genesis Panda, moved within 24 hours of disclosure, the security firm said. VulnCheck’s own experience illustrates the speed: it published 14 flaws in Copeland refrigeration controllers on Friday, two of which are under active exploitation, and saw probes hit its honeypots within minutes. Those are the same class of devices that Iranian hackers have abused to spoil food and start fires in European supermarkets.
A parallel problem is emerging inside Anthropic’s Project Glasswing, the AI-driven vulnerability discovery effort the company launched in April with $100 million in model credits and $4 million in open-source security funding, backed by Apple, Google, Microsoft, AWS and Nvidia. VulnCheck’s tracking shows Claude has produced 37,966 candidate findings, of which only 217 appear as fixed in public disclosure records. Withdrawn items number 419 and duplicates 1,654. The gap has widened quickly: nearly 12,000 candidates were added in the past two months while fixed records rose by just 15. The 442 CVEs attributed to Anthropic account for about 0.5 percent of this year’s global total.
Anthropic’s own figures differ in scale but point the same direction. As of Nov. 2, it said 6,376 of the 37,966 candidates had been reviewed by outside security experts, with 6,009 confirmed valid, 5,087 reported to maintainers and 629 patched upstream. The company has acknowledged that human review capacity is a severe bottleneck and in October introduced OSS Scanner, which lets open-source maintainers request free scans so reports reach them without waiting for Anthropic’s review queue.
The strain is heaviest on open-source maintainers. curl’s lead developer, Daniel Stenberg, received 45 reports from Anthropic in early November on top of a steady stream that he has said has overwhelmed the project’s core team. curl has since changed its disclosure rules so that low- and medium-severity reports go straight into a public GitHub repository, leaving only high and critical flaws in private handling. Of 34 reports made public so far, 19 were judged not to be vulnerabilities and six were confirmed low severity. Stenberg shut down curl’s bug bounty program earlier this year as AI-generated submissions piled up. The effort has not been fruitless: in September, Claude found a use-after-free memory error in curl that, under specific conditions, could allow a malicious server to execute remote code, which Stenberg called one of the project’s most serious flaws in years.
Bressers argued the industry must shift from chasing individual CVEs to managing risk. “We don’t have a security problem anymore. We have a data problem now,” he said, urging organizations to build infrastructure heat maps and focus on the flaws that actually affect them. He pointed to a June directive from the U.S. Cybersecurity and Infrastructure Security Agency that replaces flat deadlines with risk-based prioritization, weighing whether an asset is internet-exposed, under active exploitation and vulnerable to automated attack. “It’s telling people to figure out the vulnerabilities that actually affect you, and then put those at the top of the list, instead of just saying we’re going to fix all the criticals,” he said.
Some infrastructure providers are already adapting. The Eclipse Foundation, which runs the Open VSX Registry, launched a managed tier in April with a 99.95 percent uptime service-level agreement for commercial users including AWS, IBM, Google, Cursor and Windsurf, channeling revenue back into maintenance. Bressers said such changes draw resistance. “No one wants to be the first to jump into the alligator-filled pool,” he said, adding that the current approach must change soon. “The thing that almost scares the bejesus out of me is while 100K sounds like a lot, in a couple years we’re going to say, remember when it was only 100K?”
For investors, the shift has implications across the security sector. Vendors selling vulnerability intelligence, prioritization and patch management stand to benefit as buyers move away from severity-score-driven programs, while large software makers face rising disclosure and remediation costs. Open-source foundations, which underpin much commercial software, are being forced toward paid support models to absorb the volume. The larger risk is operational: with exploitation now measured in hours rather than months, the value of a security tool increasingly depends on how fast it can tell an organization which of tens of thousands of flaws actually threatens its systems.