Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarZDNet AI→ original

AI Finds Unknown V8 Vulnerability for $4.24 in 13 Minutes, After Failing for $350 in February

A zero-day vulnerability is a flaw the developer doesn't know about yet, so no patch exists. Xint.io gave AI models files from Chrome's V8 engine containing one such hidden bug. In February, the best models spent 25 hours and $350 and did not find it. In September, GPT 6 Sol found it in 13 minutes for $4.24. For defenders and attackers alike, it's code review that has dropped to the price of a cup of coffee. That is a reason to scan your own code with these models before attackers do.

AI-processed from ZDNet AI; edited by Hamidun News
AI Finds Unknown V8 Vulnerability for $4.24 in 13 Minutes, After Failing for $350 in February
Source: ZDNet AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceZDNet AI ↗
Editor-in-Chief
Key metric #1Claude Opus 5 found the flaw on the first try, but at a cost of $43.40
GPT 5.6 Sol in July$22, 44 minutes and two prompts
Key metric #3DeepSeek and Z.ai models did not solve the task: three of the seven found the bug

A zero-day vulnerability is a bug in software that its developer doesn't yet know about, so no fix exists for it. Hunting for such bugs with neural networks used to be expensive and futile. Pentest firm Xint.

io has shown how that has changed. In February 2026, the best models spent 25 hours and $350 and found nothing. In September, GPT 6 Sol found a critical flaw in V8, the engine that runs JavaScript in Chrome, in 13 minutes for $4.

24. Picture a code review that once produced no result after a full day and hundreds of dollars, and now costs about as much as a cup of coffee. For teams responsible for product security, that is a reason not to put off AI-based checks until next year.

How the experiment worked

The researchers took a real V8 bug, CVE-2026-102321. It is a type confusion: the program treats data of one type as another, which allows remote execution of arbitrary code. Xint employee Taisic Yun found it and reported it privately to Google's Vulnerability Reward Program (VRP). That means no information about it could have been in the models' training data, so they could not have recalled the answer from memory.

The models received a limited set of source files, including the one containing the bug, and were asked to look for security issues. The researchers tested seven models released since February 2026: two from OpenAI, three from Anthropic, and two open-weight models from DeepSeek and Z.ai. They ran the same task in different months to see how results changed over time.

Results by month

In February, the best available models scanned for 25 hours, spent $350 and did not find the critical flaw. In July, GPT 5.6 Sol succeeded in 44 minutes for $22. By September, GPT 6 Sol found the same bug in 13 minutes for $4.24, with the first guiding prompt.

Of the seven models, only three found the flaw: GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5. All three were released in July or later. GPT 5.6 Sol needed two prompts. Claude Opus 5 succeeded on the first try, but it cost $43.40, roughly ten times more than the best result. The DeepSeek and Z.ai models did not solve the task.

The backdrop is falling token prices. Tokens are the units of text you pay for when using a model. According to figures cited in a ZDNET article, they have fallen by roughly 47% per quarter on average since 2023, and by 15% in the last month alone. The cause is fierce competition among OpenAI, Anthropic and other companies. That benefits businesses, but it also makes AI-assisted attacks cheaper.

What to do about it

The report's authors argue that even "mediocre hackers" can now use frontier models to find, and potentially to exploit, vulnerabilities. Complete novices with "untamed" models are still a long way off. Experience is still needed to steer the AI correctly: without hints, the model sometimes missed the target and required extra runs, which raised time and cost. According to the researchers, hacking still requires technical literacy, but no longer the level of the industry's top specialists.

Practical steps:

  • Run the checks yourself. Your security team or contracted pentesters can run the same models over your code before attackers do.
  • Start with the most exposed areas: public-facing services and a supply chain that hasn't been audited in years.
  • Don't neglect the basics: zero-trust access, network segmentation and endpoint monitoring. These measures may deliver as much as the AI itself.

Caveats

This is a single test on a single known bug, using a trimmed-down set of files, not a search for flaws across all of Chrome's code. In a real project there are more files and more noise, and finding a flaw is not the same as writing a working exploit. In addition, the report was published by a company that sells penetration testing services.

By definition, a zero-day is unknown to the vendor, and there may be no patch at the time of an attack. AI helps find such holes faster but does not replace people: it relieves defenders of routine work and frees up time for tasks that need a human brain. The report's main conclusion is that the barrier to entry for vulnerability hunting has dropped noticeably, and that should now be factored into your threat model.

Story

Key Questions Answered

How the experiment worked?

The researchers took a real V8 bug, CVE-2026-102321. It is a type confusion: the program treats data of one type as another, which allows remote execution of arbitrary code. Xint employee Taisic Yun found it and reported it privately to Google's Vulnerability Reward Program (VRP). That means no information about it could have been in the models' training data, so they could not have recalled the answer from memory.

What are the key details on Results by month?

In February, the best available models scanned for 25 hours, spent $350 and did not find the critical flaw. In July, GPT 5.6 Sol succeeded in 44 minutes for $22. By September, GPT 6 Sol found the same bug in 13 minutes for $4.24, with the first guiding prompt. Of the seven models, only three found the flaw: GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5. All three were released in July or later. GPT 5.6 Sol needed two prompts.

What to do about it?

The report's authors argue that even "mediocre hackers" can now use frontier models to find, and potentially to exploit, vulnerabilities. Complete novices with "untamed" models are still a long way off. Experience is still needed to steer the AI correctly: without hints, the model sometimes missed the target and required extra runs, which raised time and cost. According to the researchers, hacking still requires technical literacy, but no longer the level of the industry's top specialists.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…