Xint.io study: AI models now find a Google V8 type-confusion zero-day (CVE-2026-102321) for as little as $4
A report by Xint.io shows that frontier AI models can now find a privately reported Google V8 type-confusion bug (CVE-2026-102321, leading to RCE) at very low cost. The bug was reported to the Google VRP by Xint researcher Taisic Yun, so it was not in the models' training data. Of seven models tested (two OpenAI, three Anthropic, DeepSeek and Z.ai), only GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5 found it. In February, 25 hours and $350 of scanning found nothing. In September, GPT 6 Sol found the bug in 13 minutes for $4.24. The scope was limited to a set of source files that included the vulnerable one, and some models needed extra guided prompts. The researchers conclude that expertise still matters but top-tier skill is no longer needed. This is research on AI capability, not an exploitation campaign. The article does not report exploitation in the wild, CVSS or fixed versions, and no vendor advisory was checked.
Entenda a falha
A vulnerabilidade
CVE-2026-102321 is a type-confusion bug in Google V8, the JavaScript engine, that leads to remote code execution (RCE). The articles do not say what an attacker needs (such as user interaction or configuration), give no CVSS score, and do not list affected or fixed versions. No vendor advisory was checked, so defenders should confirm patch status with Google directly.
Como funciona
The articles do not describe how the vulnerable code path works. They only say it is a type-confusion bug in V8 that can lead to RCE, located in one of the source files the researchers gave to the AI models. Nothing more about the input or the sequence of steps is public in the sources.
Por que é vulnerável
The articles give no root-cause detail beyond the bug class: type confusion, where the engine treats a value as a different type than it really is. They do not explain why it went unnoticed or what the fix changes.
Como foi descoberta
Xint researcher Taisic Yun reported the bug to the Google Vulnerability Reward Program (VRP), so it was not in the AI models' training data. Xint then used it as a target to measure how well AI models find unknown bugs. In February, 25 hours and $350 of scanning found nothing. In July, GPT 5.6 Sol found it in 44 minutes for $22. In September, GPT 6 Sol found it in 13 minutes for $4.24. Of seven models tested, only GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5 found it; scope was limited to a set of source files including the vulnerable one, and some models needed extra guided prompts.
Explicação gerada por IA a partir da matéria e dos artigos, em 16/11/2026 12:16.
Detalhes técnicos
- Flaw: type confusion in the V8 engine leading to RCE (CVE-2026-102321).
- GPT 5.6 Sol needed two prompts. Claude Opus 5 found it on its first try at $43.40.
- DeepSeek and Z.ai models failed.
Remediação
- Run AI-assisted code review and pentests on public-facing services and the supply chain.
- Maintain zero-trust access controls, network segmentation and endpoint monitoring.
Linha do tempo
-
Xint.io study: AI models now find a Google V8 type-confusion zero-day (CVE-2026-102321) for as little as $4
ler a análise desta fonte ocultar
A report by Xint.io shows that frontier AI models can now find a privately reported Google V8 type-confusion bug (CVE-2026-102321, leading to RCE) at very low cost. The bug was reported to the Google VRP by Xint researcher Taisic Yun, so it was not in the models' training data. Of seven models tested (two OpenAI, three Anthropic, DeepSeek and Z.ai), only GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5 found it. In February, 25 hours and $350 of scanning found nothing. In September, GPT 6 Sol found the bug in 13 minutes for $4.24. The scope was limited to a set of source files that included the vulnerable one, and some models needed extra guided prompts. The researchers conclude that expertise still matters but top-tier skill is no longer needed. This is research on AI capability, not an exploitation campaign. The article does not report exploitation in the wild, CVSS or fixed versions, and no vendor advisory was checked.
Technical Details
- Flaw: type confusion in the V8 engine leading to RCE (CVE-2026-102321).
- GPT 5.6 Sol needed two prompts. Claude Opus 5 found it on its first try at $43.40.
- DeepSeek and Z.ai models failed.
Remediation
- Run AI-assisted code review and pentests on public-facing services and the supply chain.
- Maintain zero-trust access controls, network segmentation and endpoint monitoring.