Osegi
Security·2 min read

AI zero-day hunting now costs $4, and 13 minutes is enough

In February the best AI models spent $350 and 25 hours without finding a Chrome engine bug. By September, one found it in 13 minutes for $4.24.

A laptop screen showing lines of code with a red warning icon, symbolizing an AI-discovered software vulnerability.
Image: ZDNET

Finding a zero-day, a security hole the software maker doesn't know about yet, used to take serious expertise and money. A new report from penetration testing firm Xint.io says an AI model can now do it in 13 minutes for $4.24.

The researchers picked a type confusion bug in Google's V8 engine (the JavaScript engine behind Chrome), tracked as CVE-2026-102321. It leads to remote code execution, meaning an attacker could run their own code. Xint researcher Taisic Yun reported it privately to Google's Vulnerability Reward Program, so it could not have been in any model's training data. Each model was pointed at a set of source files, including the vulnerable one, and told to look for security problems.

In February, the best models available spent 25 hours and $350 scanning and found nothing. In July, GPT 5.6 Sol found the bug in 44 minutes for $22. In September, GPT 6 Sol did it in 13 minutes for $4.24 with its first guided prompt.

Which AI models found the zero-day?

Xint tested seven models released since February 2026: two from OpenAI, three from Anthropic, and open-source models from DeepSeek and Z.ai. Only three found the bug: GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5, all launched from July onward. GPT 5.6 Sol needed two prompts. Claude Opus 5 got it on the first try but cost $43.40. The DeepSeek and Z.ai models failed.

Prices help explain the shift. The report cites studies showing average token costs have fallen about 47% per quarter since 2023, and 15% last month alone.

What does it mean for defenders?

The researchers say "mediocre hackers" can now use frontier models effectively, though we are "still far from complete amateurs using unharnessed models to hack systems." Expertise still matters: some models needed guidance and extra runs, which added time and cost.

The same tools work for defenders. Security teams can run the same checks first, focusing on public-facing services and supply chains that haven't been audited in years. Zero-trust access controls, network segmentation and endpoint monitoring still count for as much as AI does.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:zero-dayai securityxintgoogle v8openaianthropic

More from this day

See all →

More in: Security

See all →