AI Found Thousands of Flaws in Days. Humans Can't Patch Them Fast Enough.
Anthropic's Claude Mythos model discovered more security holes in major software than years of human review had caught. That's good news for defenders in theory, but the gap between finding a flaw and fixing it was already brutal before AI joined the hunt.
Key points
- Anthropic's Claude Mythos AI found thousands of high-severity security flaws in major operating systems and browsers in 2026, fewer than 1% of which had been fully patched at announcement.
- In 2025, security researchers published 46,407 CVEs (Common Vulnerabilities and Exposures, officially catalogued software flaws), up 16% from 40,009 in 2024.
- By 2024, attackers were exploiting new flaws on average one day before patches were even publicly available, according to a Mandiant analysis.
- The median time for criminals to exploit a known flaw is now under five days; the average time for organisations to patch a critical flaw exceeds 60 days.
- A medium-severity flaw in Palo Alto Networks firewalls, CVE-2024-9474, was chained with a critical-rated flaw to compromise roughly 2,000 devices, illustrating why severity scores alone mislead defenders.
On 7 April 2026, Anthropic announced Project Glasswing and its AI model Claude Mythos. The model did something remarkable: it trawled through major operating systems and web browsers and surfaced thousands of serious security holes, many of which had sat undetected for decades despite human review and automated scanning. Fewer than one in a hundred of those flaws had been fully patched by the time the announcement landed.
Why does an AI finding more bugs make things worse?
More discovered flaws means more work for the people whose job it is to fix software, and that workforce was already drowning before Mythos arrived.
Even without superhuman AI assistance, the security community logged 46,407 CVEs in 2025, a 16% jump from the year before. The US National Institute of Standards and Technology processed nearly 42,000 of them, 45% more than any prior year, and still fell behind. Every new flaw Mythos surfaces will eventually reach company security teams as a fresh item on an already unmanageable list.
The patching side hasn't kept pace, and the numbers show how badly.
| Year | Median days from disclosure to first exploit |
|---|---|
| 2018 | 771 days |
| 2021 | 84 days |
| 2023 | 6 days |
| 2024 | Under 5 days (average: -1 day) |
That final row deserves a plain-English read: in 2024, Mandiant researchers found attackers were, on average, exploiting flaws before patches were publicly released. Criminals watching the same code repositories as defenders are now moving faster than the official disclosure system. Our coverage on 7 August 2026 of the 1Password study on AI patch quality adds another wrinkle: even when AI tools do attempt fixes, they get them right less than a third of the time.
Is the traditional way of deciding what to patch still working?
No, and a real-world example shows exactly why.
Most security teams rank flaws using CVSS scores (the Common Vulnerability Scoring System, a number from 1 to 10 estimating how dangerous a flaw is in isolation). Patch the 10s first, then the 9s, work down the list. Feels orderly. But in 2025, 28% of the flaws attackers actually exploited carried medium CVSS scores, meaning more than a quarter of real attacks came from the middle of the queue teams were deprioritising. Chris Gibson, executive director of FIRST, said organisations relying on CVSS base scores alone are "the least apt and accurate" at prioritisation, in comments reported by Dark Reading.
The Palo Alto Networks incident makes the problem concrete. CVE-2024-9474, scored 6.9 (medium), was combined with CVE-2024-0012, scored 9.3 (critical), to compromise roughly 2,000 firewalls. A score-first team would have patched the critical flaw and moved on. Attackers saw one connected route and used both.
What should organisations actually do differently?
Patch strategically, not just urgently. Security teams are increasingly advised to model attack paths the way a burglar cases a building: not "which window is easiest to break?" but "which route from the window reaches the safe?"
In practice that means asking whether an internet-facing flaw can actually reach a sensitive database, then identifying the single points where fixing one vulnerability disconnects the most possible attack routes, and patching those choke points first, even when their CVSS score looks unimpressive.
A medium-severity credential flaw sitting at the crossroads of every path to a company's core servers matters more than a critical flaw on a machine that can't talk to anything else on the network. That's the whole argument, distilled. The analyst community, including the Dark Reading opinion piece that prompted this story, calls it thinking in chains rather than checklists, and it's the right frame.
For ordinary people, the message is less dramatic but still real: when your employer pushes a software update, it's increasingly racing a clock measured in hours. Apply it.
Common questions
Does this affect the software I use at home?
Potentially yes. When Mythos or similar tools find flaws in widely used operating systems and browsers, patches will eventually reach consumer devices. Keeping your phone and laptop updated is the most useful thing you can do.
Should I be worried that AI is being used to find flaws in software?
The security community hopes the same AI tools will help defenders fix flaws faster than attackers exploit them. For now, evidence suggests both sides are picking up the technology and defenders are still catching up.
