Sean Heelan ran an experiment where LLM agents built on Opus 4.5 and GPT-5.2 wrote over 40 exploits for a zero-day in the QuickJS JavaScript interpreter. The agents succeeded across six scenarios, handling modern mitigations like ASLR and shadow stacks, with GPT-5.2 solving every task. The code and full results are available on GitHub.

The core conclusion is that exploit development is heading toward industrialisation. The limiting factor for future attacks may not be the number of hackers, but an organization's token throughput. The agents operated autonomously, reading source code and debugging, turning the vulnerability into an API to read and modify process memory.

Verification was straightforward: an exploit works if it achieves its goal, like spawning a shell or writing a file. However, not all intrusion tasks are easily industrialized. Initial access and lateral movement require real-time interaction, where a single wrong action can terminate the search. This creates a harder problem for current models than offline exploit generation.