Zero-Day Vulnerabilities: Xint's $4.24 Test Remains Steered

According to Xint.io, GPT 6 Sol identified a vulnerability in Google V8 in 13 minutes, for $4.24, based on a guided prompt. This performance does not eliminate the need for technical guidance or a prepared research scope.
Xint Still Believes Guidance for Models is Necessary
Access to these capabilities does not mean that any amateur can conduct an intrusion. Xint researchers believe that technical skills remain essential, even if highly advanced expertise is no longer systematically required. They consider the scenario of amateurs hacking applications or systems with models they do not master to be far off.
According to Xint, knowing how to guide a model remains crucial for identifying a vulnerability. Some trials required assistance and multiple attempts, resulting in additional costs and extended work. Nevertheless, researchers believe that technical advancements, combined with decreasing prices, make research and even potential exploits accessible to mid-level hackers that were previously too difficult or expensive.
Three Models Find the Flaw in a Limited Scope
Xint.io compared seven models released since February 2026: two from OpenAI, three from Anthropic, as well as the open-source models from DeepSeek and Z.ai. The team tracked their results on an identical target to measure the evolution of their capabilities over the past year. They report that some tasks that were missed a few months ago are now successfully completed.
The target was CVE-2026-102321, a type confusion affecting Google V8 that allows for remote code execution. Taisic Yun, a researcher at Xint, had confidentially submitted it to Google's reward program; according to Xint, it was absent from the data used to train the models. The systems received a selection of source files, including the one containing the flaw, with instructions to look for security issues, without it concerning an entire software.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
In February, Xint reported a failure of the best available models despite 25 hours of research and $350 spent. In July, GPT 5.6 Sol identified the flaw after two prompts, in 44 minutes, for $22. The result from September with GPT 6 Sol, mentioned in the lead, was achieved on the first guided prompt.
Claude Opus 5 also succeeded on its first attempt, with a bill of $43.40, according to Xint. These models are the only ones that succeeded; all were launched starting in July. The systems from DeepSeek and Z.ai failed.
Price Reductions Also Benefit Security Teams
Major developers are seeking to make their models more efficient, while Xint observes progress in the detection of zero-day vulnerabilities. Studies estimate the average decrease in token costs to be around 47% per quarter since 2023, depending on the domains, and 15% last month. This pressure on prices is attributed to competition, particularly between OpenAI and Anthropic. It can benefit companies developing services or improving their productivity, but it also makes attacks using AI against unknown flaws less costly.
Defenders can also benefit from faster, more competent, and cheaper models. Internal teams like pentesters can conduct similar research before attackers do. For organizations, the recommended adaptation includes engaging in AI-assisted testing without waiting for the next budget review. A risk assessment targeting publicly accessible services or digital supply chains that have long gone without audit can reduce exposure to AI-facilitated attacks.
Since a zero-day flaw is unknown to the provider, no patch is necessarily available at the time of its revelation or exploitation. AI can free up time for tasks requiring human intervention, without replacing all protections. Access governed by a zero-trust approach, network segmentation and hardening, as well as monitoring for suspicious behavior on endpoints can offer comparable, if not superior, effectiveness.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.