AI Model Builds a Working Chrome Exploit in Three Days, Researcher Says Exploit Development Has Changed

A security researcher at Hacktron has reported that a new frontier AI model, GPT-5.6 Sol Ultra, was able to build a full Chrome renderer exploit from start to finish. The test used a recent Chrome release (version 149.0.7827.201) and asked the models to study published security-fix commits in its V8 JavaScript engine. After three days, the researcher says Sol Ultra succeeded in launching the calculator app on the test machine with only limited nudges. The researcher describes this as a huge jump from an earlier test three months ago with a different model, and argues that exploit development as we know it is dead for anyone who can apply large-scale computing power to a model of this level or above.

The researcher reports that the model relied heavily on helper agents it created itself. The main agent used about 819.6 million input tokens, while its 74 sub-agents used about 1.27 billion, meaning roughly 70% of the investigation happened in sub-agents. The model also repeatedly condensed its own working notes without losing track of the task. Two other models, Grok 4.5 and GPT-5.6 Sol Medium, got as far as leaking information but stalled and were stopped.

The flaw involved was a bug in how Chrome's Maglev compiler handled certain array operations, and a patch has already been released. This is the researcher's own account, and the excerpt we reviewed is only the opening of the article, so readers should consult the original for full detail.