AI Models Can Discover Zero‑Day Vulnerabilities for as Little as $4, Threatening Wider Exploitability
What Happened — A new report from penetration‑testing firm Xint.io shows that several frontier large‑language models (LLMs) can locate a critical, previously unknown vulnerability in Google’s V8 JavaScript engine (CVE‑2026‑102321) for a cost as low as $4.24 and in under 15 minutes. The study evaluated seven models released between February 2026 and September 2026 – two from OpenAI, three from Anthropic, and two open‑source offerings from DeepSeek and Z.ai. Only three models succeeded: GPT‑6 Sol (found the bug in 13 minutes for $4.24), GPT‑5.6 Sol (44 minutes for $22), and Claude Opus 5 (first‑try success at $43.40). Earlier tests in February required 25 hours and $350 without success, illustrating rapid improvement in AI‑driven vulnerability discovery.
The researchers fed each model a limited set of source files containing the V8 bug and prompted them to “search for security issues.” The successful models required minimal guidance, indicating that even “mediocre hackers” could leverage these tools without deep expertise. Xint.io warns that falling token prices – down ~47 % per quarter since 2023 – make such attacks financially feasible for a broader threat pool.
Zero‑day flaws, by definition, are unknown to vendors and lack patches when disclosed. While the V8 bug was responsibly reported to Google’s Vulnerability Reward Program, the study demonstrates that AI can accelerate the discovery phase, shrinking the window between vulnerability emergence and potential exploitation.
The report also notes that AI‑assisted testing is not a silver bullet. Models sometimes failed without precise prompts, and expertise remains essential to guide the AI and interpret results. Nonetheless, the speed and cost advantage reshapes the attacker’s cost‑benefit calculus, pushing organizations to treat AI‑generated findings as a realistic threat vector.
Security teams are urged to incorporate AI‑driven scanning into their vulnerability management programs, prioritize rapid patching of high‑severity findings, and reinforce zero‑trust architectures to limit the impact of any undisclosed flaw that may be weaponized.
Why It Matters for Trust & Control Assurance — This development tests the control objective of continuous vulnerability discovery and timely remediation, underscoring the need for automated, evidence‑driven processes.
Who Is Affected — Technology and SaaS providers, software vendors, financial services firms, and any organization exposing public‑facing applications or digital supply‑chain components.
Recommended Actions
- Integrate AI‑assisted scanning tools into your vulnerability management workflow and treat findings as audit‑ready evidence.
- Map AI‑generated results to your existing control framework (e.g., NIST CSF Identify → Risk Assessment) and document remediation timelines.
- Strengthen zero‑trust network segmentation and endpoint monitoring to contain potential exploitation of undisclosed flaws.
Source: ZDNet article
Technical Notes
- Targeted vulnerability: CVE‑2026‑102321, a type‑confusion bug in Google V8 that enables remote code execution.
- Models evaluated: GPT‑6 Sol, GPT‑5.6 Sol, Claude Opus 5 (successful); other models (DeepSeek, Z.ai) failed.
- Cost per successful discovery ranged from $4.24 to $43.40; time to discovery from 13 minutes to 44 minutes.
- Token price decline (~47 % per quarter since 2023) drives affordability.
Source: ZDNet article

