Frontier AI Models Identify Zero-Day Flaws in Minutes for Under $5, Study Shows
Research from Xint.io shows how falling token prices and advancing AI models allow attackers and defenders to discover zero-day bugs in minutes for under $5.

The rapidly falling cost of compute combined with rapid advances in frontier AI models has dramatically lowered the economic and technical barriers to discovering zero-day vulnerabilities. Research released by cybersecurity firm Xint.io demonstrates that finding critical software flaws—a task that previously required specialized security expertise and hundreds of dollars in API credits—can now be accomplished by top-tier models in minutes for less than $5.
The study examined seven flagship AI models released since February 2026, including offerings from OpenAI, Anthropic, DeepSeek, and Z.ai. To evaluate their security capabilities, researchers tested the models against a real-world type confusion bug in Google’s V8 engine. The vulnerability, tracked as CVE-2026-102321, allows remote code execution (RCE) and was privately reported to the Google Vulnerability Reward Program by Xint researcher Taisic Yun, ensuring it was not present in any model's training data.
How AI vulnerability discovery evolved in 2026
The performance leaps documented in the report occurred over a span of just seven months. In initial testing conducted in February 2026, state-of-the-art models spent 25 hours scanning the target source files at a total API cost of $350 without successfully identifying the bug.
By July, performance shifted noticeably. OpenAI's GPT 5.6 Sol uncovered the zero-day flaw in 44 minutes across two prompts, incurring a cost of $22. In September, GPT 6 Sol successfully pinpointed the exact vulnerability in just 13 minutes on its first guided prompt, costing only $4.24.
Out of the seven models tested, only three were capable of finding the bug: GPT 6 Sol, GPT 5.6 Sol, and Anthropic’s Claude Opus 5—all of which were launched in or after July 2026. While Claude Opus 5 successfully located the flaw on its first attempt, it did so at a higher cost of $43.40. Meanwhile, open-source models from DeepSeek and Z.ai failed to identify the security flaw altogether.
This sharp drop in discovery costs reflects broader market trends. Driven by aggressive competition among major developers like OpenAI and Anthropic, token prices across the industry have decreased by approximately 47% per quarter since 2023, with prices dropping another 15% in the last month alone.
What it means for developers
The plummeting cost of model execution paired with heightened reasoning capabilities reshapes the threat landscape for software engineers and maintainers. According to Xint researchers, identifying high-severity zero-days no longer demands top-tier domain expertise. While complete novices cannot yet deploy unharnessed models to breach systems, attackers with intermediate technical skills can leverage frontier models to discover actionable vulnerabilities that were previously too complex or expensive to pursue.
However, expertise remains a critical factor in successfully guiding these AI tools. In testing, models required tailored prompts and constrained source scopes to locate the bug, meaning prompt quality directly affects discovery speed and overall token expenditure.
For development and security teams, these same automated inspection techniques can be integrated into regular auditing pipelines to catch flaws before external threat actors do. To run automated audits, automated vulnerability scanning, or build security tools cost-effectively, developers can access top AI models like Claude, GPT, and DeepSeek through a single API key at https://apixoai.online.
Defending systems in an era of cheap AI analysis
Because zero-day vulnerabilities are unknown to vendors until disclosed or exploited, relying solely on reactive patching is no longer sufficient. Security teams are encouraged to prioritize risk assessments on high-value attack surfaces, such as public-facing services and legacy digital supply chain dependencies.
Furthermore, defensive strategies must extend beyond AI-driven scanning. Human oversight remains indispensable, as AI models can free up security teams to focus on nuanced architecture reviews and complex problem-solving. Implementing zero-trust access controls, enforcing strict network segmentation, and continuously monitoring endpoints for anomalous behavior remain foundational defenses against AI-assisted exploits.
Source: AI finds zero-day flaws for just $4 now – even mediocre hackers can cash in — ZDNet AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key

