Skip to content
Qubitron Security

4 min readQubitron Labs security team

AI now finds vulnerabilities at volume. What changes for your patch cycle

AI labs, Google, VulnCheck, five national agencies and CERT-In all weighed in this year. The honest read: discovery has sped up, mass exploitation hasn't followed yet.

  • AI security
  • Penetration testing
  • Vulnerability management

Two kinds of story about AI and security ran side by side in 2026. One says AI agents now find and exploit vulnerabilities faster than any team can patch. The other says the impact so far is real but modest. Both are backed by data, and both are worth reading closely, because the gap between them is where your patch policy lives.

What the labs and contests show

  • Autonomous discovery works at scale. In DARPA's AI Cyber Challenge final in August 2025, competing systems found 54 synthetic vulnerabilities and patched 43, and also found 18 real ones across more than 54 million lines of code. A patch took about 45 minutes on average.
  • It has reached bug bounties. XBOW reported in June 2025 that its autonomous pentester reached the top of HackerOne's US leaderboard, having submitted nearly 1,060 reports. Its own security team reviewed them before submission.
  • Frontier models exploit, not just find. In April 2026, Anthropic wrote that Claude Mythos Preview can identify and exploit zero-days in every major operating system and browser, "when directed by a user to do so". In one browser-engine test it produced working exploits 181 times, against twice for the previous model. At publication, Anthropic said fewer than 1% of the potential vulnerabilities it had found were fully patched.
  • Attackers use it already. In November 2025, Anthropic described disrupting a campaign it attributes with high confidence to a Chinese state-sponsored group, which used Claude Code against about thirty targets. AI performed most of the work, and the campaign succeeded in a small number of cases. The model also "occasionally hallucinated credentials".

What governments are telling defenders

  • The Five Eyes. In June 2026, the heads of the cyber agencies of the UK, US (NSA and CISA), Australia, Canada and New Zealand said AI "accelerates the speed, scale, and sophistication of cyber threats", and that the timeline is "not years, it is months".
  • NCSC UK. Its CTO told organisations in May 2026 to prepare for a vulnerability patch wave, because skilled people using AI are showing they can exploit accumulated technical debt at scale.
  • CERT-In. India's high-severity advisory CIAD-2026-0020, issued in April 2026, warns that frontier AI may be able to find vulnerabilities autonomously and chain multi-stage attacks.
  • CERT-EU. Citing Google's M-Trends 2026, CERT-EU wrote that the mean time to exploit new vulnerabilities has fallen to an estimated negative seven days: exploitation now often comes before disclosure. Typical enterprise patch cycles still run 30 to 60 days.

The counterweight

Two careful datasets temper the alarm.

  • Google Threat Intelligence Group found that half of AI-discovered vulnerabilities lead to remote code execution, against 26% across all CVEs, though it says that probably reflects how research programs aim these systems. It saw one threat cluster exploit an AI-found flaw within four days of disclosure. Yet only 0.23% of all 2026 disclosures, roughly 1 in 431, were observed exploited.
  • VulnCheck counted 1,061 vulnerabilities attributed to AI-assisted discovery in the first half of 2026, of which 14 (1.3%) were confirmed exploited in the wild. Its verdict: "the impact has been real but modest."

So discovery has clearly sped up. Mass exploitation of what is discovered has not, yet. Most headline numbers also come from the labs and vendors themselves.

What this changes for you

  • Shorten the patch window for what faces the internet. If exploitation can come before disclosure, 30 to 60 days is too slow for edge devices, auth services and public APIs. Triage those first.
  • Know what you run. You can't patch fast what you can't list. Keep an inventory of internet-facing services and their versions.
  • Expect more reports, not just more attacks. If you run a disclosure program or bug bounty, plan for volume and for duplicate, AI-written reports.
  • Test the logic, not only the versions. AI tooling is very good at known bug classes in code it can read. Business-logic flaws, authorization gaps and chained issues in your own app still need a person who understands what the app is for.
  • Insist on verified findings. Whoever tests your app, AI-assisted or not, every reported issue should come with reproduction steps and evidence.

That last point is how we run a penetration test: AI-assisted discovery to cover ground, and a person who reproduces every finding before it reaches you.

Sources

Keep reading

All posts

Next step

Find out what your app gives away. Book a free exposure call.

Thirty minutes with an engineer, not a sales deck. Tell us what you ship; we tell you what we'd look at first and what an audit would involve.

Book a free exposure callRequest an audit

We only test applications you own or are explicitly authorized to test.