Skip to main content

AI model found a V8 flaw for $4.24 in one test

17 NOVEMBER 2026·2 MIN READ·1 SOURCE

A Xint report says GPT 6 Sol found a vulnerability in Google’s V8 engine in 13 minutes for $4.24 in September. The result came from a test of seven AI models and a limited set of source files that included the flaw.

AI model found a V8 flaw for $4.24 in one test

Key takeaways · 4

  • 01

    GPT 6 Sol found the flaw in 13 minutes for $4.24 after its first guided prompt.

  • 02

    GPT 5.6 Sol and Claude Opus 5 also found the flaw; Claude Opus 5 did so on its first attempt for $43.40.

  • 03

    The test used limited source files, including the file containing the flaw, so the results do not establish how models perform with no such access.

  • 04

    Xint recommends that organizations have security teams or penetration testers check systems, prioritizing unaudited public-facing services and digital supply chains.

A targeted test of seven models

The report examined seven models launched since February 2026: two from OpenAI, three from Anthropic, and one each from DeepSeek and Z.ai.[1] Researchers provided a limited set of source files, including the file containing the vulnerability, and asked the models to find a type-confusion flaw in Google’s V8 engine.[1] The flaw, tracked as CVE-2026-102321, could enable remote code execution and was disclosed to Google’s Vulnerability Reward Program by Xint researcher Taisic Yun.[1] Xint said the vulnerability was not in the models’ training data.[1]

Costs and results changed over time

In February, the best models tested did not find the flaw after 25 hours of scanning that cost $350.[1] In July, GPT 5.6 Sol found it in 44 minutes for $22.[1] By September, GPT 6 Sol found it in 13 minutes for $4.24 using its first guided prompt.[1] Three models succeeded overall: GPT 6 Sol, GPT 5.6 Sol, and Claude Opus 5.[1] Claude Opus 5 found the vulnerability on its first attempt, at a cost of $43.40.[1] The DeepSeek and Z.ai models did not find it.[1]

The low price has important limits

Xint attributed the declining search cost to model improvements and falling token prices.[1] Studies cited in its report estimate that average token costs have fallen about 47% per quarter since 2023, and 15% in the last month.[1] Xint attributed lower token prices to competition among companies including OpenAI and Anthropic.[1] The test was not a demonstration of effortless hacking: in some cases, models could not identify the target flaw without guidance, and additional runs increased time and cost.[1] Xint said that inexperienced people using unmodified models to hack systems remain a distant prospect.[1]

What defenders should consider

Xint said mediocre hackers can use frontier models to find, and potentially exploit, vulnerabilities.[1] It recommends that organizations have their security teams or penetration testers run similar checks before attackers do.[1] Xint also recommends prioritizing risk assessments for public-facing services and digital supply chains that have gone unaudited for years.[1] The report notes that zero-day flaws are unknown to vendors and may have no patch when disclosed or actively exploited.[1] Xint says zero-trust access controls, network segmentation, stronger network defenses, and endpoint monitoring can help protect against threats.[1] It argues that AI could free defenders to focus on work requiring human judgment.[1]

The results make it worth reassessing how much security testing organizations can run, while the limited-file setup and need for guidance caution against treating the reported cost as a universal benchmark. Professionals can use the findings to focus review efforts on exposed or unaudited systems and consider where human judgment remains essential.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 17 November 2026

    AI model found a V8 flaw for $4.24 in one test

Sources

AI fluency, one session a day, built for your work.