Armin Ronacher, the developer behind the Flask web framework, summed up 2026 in security in one sentence: “As fun as AI is, we built so much bad software over the last 30 years, and now the universe is punishing us for our sins.” He’s right. AI models aren’t writing many new bugs into old software. They’re finding the ones we left there decades ago, faster than anyone can fix them.
The line was posted on X and picked up by Simon Willison on 22 October 2026. The numbers behind it are worth a look.
Key takeaways
- AI vulnerability hunters such as Anthropic’s Mythos and OpenAI’s GPT-5.4-Cyber are surfacing flaws in old code at record speed.
- The US National Vulnerability Database logged 45,207 flaws between 1 January and 27 July 2026, close to the whole of 2025.
- FIRST, the global body for incident response teams, expects about 66,000 CVEs in 2026.
- Finding bugs is no longer the hard part. Triaging and patching them is.
- So far, the number of flaws actively exploited in the wild has not risen in step.
What did Armin Ronacher actually mean?
The “sins” are the shortcuts the industry has taken since the 1990s: memory-unsafe C code, parsers nobody audited, features shipped on deadline. For years those bugs stayed hidden because finding them took a skilled human weeks of work.
AI changed the economics. A model can read a codebase, form a hunch, write a test and try again, thousands of times, without getting bored. Think of a house whose damp nobody spotted for 30 years, and then someone turns on a thermal camera. The rot was always there. Now you can see all of it at once, and the bill comes due in one go.
Ronacher’s choice of words matters too. He isn’t blaming AI. He’s blaming us.
How big is the AI-driven bug flood?
Big enough to break records at the largest software companies. Bloomberg reporting republished by The Star on 28 July 2026 laid out the jump:
| Vendor or tracker | 2026 figure | Comparison |
|---|---|---|
| Oracle, July patch update | 1,449 fixes | 309 in the same update a year earlier |
| Microsoft, July disclosures | 642 bugs | Nearly 5x July 2025 |
| Google Chrome, one update | 433 fixes | 11 in the equivalent update a year earlier |
| National Vulnerability Database | 45,207 flaws (1 Jan to 27 Jul) | Close to the full 2025 record |
IBM’s X-Force team found the same pattern. In its analysis of AI vulnerability discovery, published CVEs (Common Vulnerabilities and Exposures, the public ID numbers given to security flaws) for Microsoft, Oracle, Chrome and Apache rose from about 1,493 in January to September 2025 to more than 10,124 over the same months of 2026. Volumes stayed close to normal through the first quarter, then climbed steeply from April.
These aren’t trivial bugs, either. Of 223 AI-attributed flaws IBM examined, 55.2% were rated High or Critical, 91% could be reached remotely and about a quarter allowed code execution.
Why the real bottleneck is people, not models
Every bug report still needs a human to confirm it, write a fix, coordinate disclosure and ship a release. That pipeline hasn’t sped up. Help Net Security reported on 15 June 2026 that FIRST expects maintainers to face roughly double their usual workload, and that a dip in published counts may reflect staff shortages rather than safer code.
Mozilla shows what the upside looks like with enough hands. A harness built on its fuzzing setup, using Anthropic’s Mythos Preview and Claude Opus 4.6 through Project Glasswing, found and fixed 271 bugs for the Firefox 150 release.
Smaller projects don’t have Mozilla’s staff. You can see the strain in this autumn’s run of emergency releases, from the PHP2Shell flaws that took websites offline to Node.js 24.22.1’s batch of eight security fixes.
Are attackers winning because of this?
Not yet, and that’s the surprising part. According to the same Bloomberg report, the US government’s Known Exploited Vulnerabilities catalogue showed no rise in actively exploited flaws in 2026. Dustin Childs of Trend Micro put it bluntly: “We just aren’t seeing the numbers to back up the doom and gloom prophets.”
The catch is speed. Alexander Leslie of Recorded Future said the average time for attackers to exploit a known flaw fell from 72 hours in 2025 to 24 hours in 2026. So fewer bugs may be weaponised than people feared, but the ones that are get used much faster.
My take: Ronacher’s line lands because it reframes the panic. This is a debt being called in, and the tool calling it in works for defenders as well as attackers.
What should you do about it?
- Shorten your patch cycle. If exploits arrive in about a day, monthly patching is too slow for anything exposed to the internet.
- Prioritise by risk. Only a small, steady share of CVEs is exploited in the wild. Track those first.
- Update the toolchain too. Security fixes often land in package managers and runtimes, as with the Python 3.15.1 release that also needed a pip update.
- Know what you run. Throwaway AI-built apps can carry flaws that never get a CVE. Keep an inventory.
What happens next
Expect the rest of 2026 to finish near FIRST’s 66,000-CVE forecast. Expect record patch bundles from big vendors and continued strain on volunteer maintainers. Vendors are also turning AI on the triage problem. Microsoft released MAI-Cyber-1-Flash on 27 July 2026 to help teams manage vulnerabilities.
The bigger question is whether the flood eventually drains. Old code has a finite number of bugs. If AI keeps finding them and teams keep fixing them, software in 2028 could be cleaner than it has ever been. That only happens if the fixing keeps up with the finding.
FAQ
Who is Armin Ronacher?
He is an Austrian software developer best known for creating Flask, the Python web framework. He writes widely about AI-assisted programming on his blog and social media.
Is AI creating more security bugs?
Mostly, AI is finding existing bugs faster. Some of the rise in CVE counts also comes from bookkeeping, as GitHub and VulnCheck have backfilled older records.
Are more flaws being exploited in 2026?
Not according to the US Known Exploited Vulnerabilities catalogue as of July 2026. Exploitation has become faster, though, with the average time dropping to about 24 hours.
Your next step: check when your most exposed system was last patched. If it was more than a week ago, start there.
2 comments