Safety & Security

AI found a Google zero-day for $4.24 in 13 minutes in 2026

Xint.io tested seven AI models against a privately disclosed Google V8 bug. Three succeeded, with GPT 6 Sol finding it in 13 minutes for $4.24 by September.

AI finds zero-day flaws for just $4 now – even mediocre hackers can cash in
AI finds zero-day flaws for just $4 now – even mediocre hackers can cash inAI-generated
By Sophie Lindqvist5 min read

Why it matters

  • Only three of seven models found the bug: GPT 6 Sol, GPT 5.6 Sol, and Claude Opus 5.
  • In February 2026, the best available models failed after 25 hours of scanning and $350 in costs.
  • GPT 5.6 Sol found the bug in 44 minutes for $22 in July 2026.
  • GPT 6 Sol found it in 13 minutes for $4.24 with its first guided prompt by September.
  • ZDNET reported that average token prices fell about 47% per quarter since 2023 and 15% last month.

GPT 6 Sol found a privately disclosed Google V8 zero-day vulnerability in 13 minutes for $4.24 in September 2026, according to an Xint.io report that tested seven AI models against the same bug. The run used its first guided prompt.

Xint.io’s best available models had needed 25 hours and $350 of scanning in February 2026 without finding the critical vulnerability. For defenders, that shift changes the economics of finding flaws before threat actors do.

What did the test cover?

The target in Xint.io’s test was CVE-2026-102321, a type-confusion bug in Google’s V8 engine that could lead to remote code execution, or RCE. Xint researcher Taisic Yun disclosed it through the Google Vulnerability Reward Program, or VRP. Xint.io said the private disclosure kept the flaw out of the models’ training datasets.

Across seven models launched since February 2026, the team tracked each model’s performance over time. The lineup included two OpenAI models, three Anthropic models, and two open-source models from DeepSeek and Z.ai. Xint.io said popular models had evolved over the past year to the point where tasks that failed only months earlier now succeed.

Xint.io limited the test to a set of source files, including the file that contained the vulnerability. Each model was told to search those files for security issues. The result is meaningful but narrow: it measures guided code review, not unaided attacks against live systems.

Why do falling token prices matter?

ZDNET, citing studies, reported that the average cost of tokens had fallen about 47% per quarter since 2023 and 15% in the previous month alone. The report said costs vary by domain and attributed the decline to intense competition among companies including OpenAI and Anthropic. It also said major AI developers were moving quickly to improve model effectiveness.

ZDNET described falling token prices as good news for businesses seeking productivity and new services from AI. It also said the trend makes AI-assisted attacks cheaper, including software breaches based on exploiting unknown vulnerabilities.

The Xint.io researchers said: “mediocre hackers” can now use frontier models effectively to discover and potentially exploit vulnerabilities. They added: “while we are still far from complete amateurs using unharnessed models to hack systems and applications, what used to be considered too hard and costly to be worth the effort of the average hacker is now reachable.”

The same economics can help defenders because lower prices make AI-assisted checks easier to run. Xint.io said organizations should adjust their security posture as model skills and speed improve. Falling token costs widen access to the capability, which raises the cost of waiting.

How did the results change?

Xint.io’s timeline separates three points in the field’s progress.

  • February 2026: After 25 hours of scanning and $350 in costs, the best models available at the time failed to find the critical vulnerability.
  • July 2026: GPT 5.6 Sol found the zero-day flaw in 44 minutes for $22.
  • September 2026: GPT 6 Sol found it in 13 minutes for $4.24 with its first guided prompt.

Only three of the seven models found the bug: GPT 6 Sol, GPT 5.6 Sol, and Claude Opus 5. All three had launched from July onward. GPT 5.6 Sol needed two prompts, while Claude Opus 5 found the bug on its first try but cost $43.40.

The DeepSeek and Z.ai models failed the task. GPT 6 Sol produced the cheapest successful run, and that run still used a guided prompt.

Can defenders run the same checks?

Xint.io says an organization’s security team or its pentesters can run the same checks and find vulnerabilities before threat actors do. The report argued that AI-assisted tests are no longer something organizations can push into next year’s budget review. That makes current security review, rather than future procurement, the relevant starting point.

The report recommends starting with a risk assessment aimed at likely weak spots. Its examples include public-facing services and a digital supply chain that has not been audited for years. The report says this focus can mitigate the risk of AI-enabled cyberattacks.

Xint.io notes that zero-day vulnerabilities are unknown to vendors, and a patch might not be available when they become public or are actively exploited in the wild. It said AI can only go so far, but it can free defenders for work that needs human judgment. Running the same checks before an attacker does turns lower model costs into a defensive advantage.

Where do human controls still matter?

Xint.io did not put AI ahead of conventional controls. It said zero-trust access controls, network segmentation, stronger network defenses, and endpoint monitoring for suspicious behavior can do as much as, or more than, AI.

That balance leaves human defenders with work AI cannot fully handle. Xint.io said expertise still matters when instructing models to discover vulnerabilities, and some models needed guidance and additional runs. Those extra runs increased time and cost, which explains why the 13-minute result should not be read as a universal rate.

The researchers concluded: “Software breaches still require some level of technical proficiency, but no longer top-of-the-distribution expertise.” AI lowers the skill floor for vulnerability hunting, but it does not remove expert judgment or routine security controls.

What can organizations conclude?

Xint.io tested one flaw within selected source files, using instructions aimed at that code. Its results show that guided code review can surface a critical zero-day in minutes, but they do not establish the same success rate for other codebases or broader security tasks. The seven-model sample, single target, and bounded file set make this a benchmark, not a general guarantee.

Three points stand out in Xint.io’s results.

  • GPT 6 Sol completed the task in 13 minutes for $4.24.
  • The three successful models were GPT 6 Sol, GPT 5.6 Sol, and Claude Opus 5; all launched from July onward.
  • Guidance remained material: GPT 5.6 Sol needed two prompts, while extra runs for other models increased time and cost.

For defenders, the source’s next step is clear: use AI-assisted checks now, starting with public-facing services and digital supply chains not audited for years. Human review and conventional controls must stay in place. As token prices fall and model skills improve, the practical warning is simple: do not postpone these checks when current models can find critical flaws in minutes.

Original: xint.io

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. GPT-6 Cyber scores 48.5% on ExploitGym, targets Linux and Firefox
  2. GPT-6 Cyber found 81 unknown Firefox flaws and 948 kernel exploits
  3. OpenAI models crossed cyber-test boundaries in two disclosed incidents
  4. Google Reports First Autonomous AI Attack That Finds and Exploits Flaws
  5. OpenAI Cancels GPT-6.1 Astra Release Over Deceptive Behavior in Testing

« Previous articleNext article »