New Study Finds GPT-5.2 Can Reliably Develop Zero-Day Exploits at Scale

Recent advancements in large language models have enabled the autonomous development of working exploits for zero-day vulnerabilities, indicating a notable shift in offensive cybersecurity strategies.

AI Capabilities in Exploit Development

Artificial intelligence systems can now perform complex exploit development tasks that traditionally required specialized human expertise. AI agents were tested to develop exploits under realistic conditions, such as modern security mitigations and unknown heap states, without relying on hardcoded memory offsets.

In six different scenarios, AI agents produced over 40 distinct working exploits. The GPT-5.2 model successfully solved every scenario, while the Opus 4.5 model solved all but two.

Security researcher Sean Heelan conducted controlled experiments using AI agents based on Anthropic’s Opus 4.5 and OpenAI’s GPT-5.2 against a previously unidentified vulnerability in the QuickJS JavaScript interpreter. The AI agents transformed the vulnerability into a functional API for reading and modifying the target process memory space, analyzing source code, debugging, and iterating through trial-and-error processes autonomously.

Most challenges were completed in under one hour with relatively low costs, with a typical agent run consuming approximately 30 million tokens at a cost of around $30 USD for Opus 4.5.

The most challenging scenario required GPT-5.2 to write a specified string to disk under active enterprise-grade protections like address space layout randomization, non-executable memory regions, and more. The AI agent developed a solution using seven function calls in glibc’s exit handler to bypass these defenses. This exploit required 50 million tokens and took three hours to develop, costing approximately $50 for that run.

The experiments highlighted two limitations: QuickJS has less complexity than production browser engines, and the exploits used known gaps and implementation flaws, similar to human exploit developers.

This research indicates a potential "industrialization" of offensive operations in cybersecurity, where computational token throughput might limit hacking capabilities instead of skilled personnel availability. Exploit development is identified as an ideal use case for AI automation due to clear verification methods and well-defined solution spaces.

The experimental code, technical write-ups, and agent outputs are publicly available on GitHub for verification and reproduction. The researcher encourages the security community to test AI capabilities against real targets rather than relying solely on synthetic datasets.

Based on reporting by GBHackers.