AI Model Builds a Working Chrome Exploit in Three Days, Raising the Stakes for Patching

Security firm Hacktron has published research describing how a frontier AI model, GPT-5.6 Sol Ultra, developed a full exploit for a recent version of Google Chrome. The researcher tested it against GPT-5.6 Sol Medium and Grok 4.5. The models were given the source code for Chrome's V8 engine and a test build, then asked to study published security-fix commits and work through three stages: build basic memory-access tools, escape Chrome's sandbox, and run an arbitrary command. Sol Ultra completed the chain and launched the calculator app with limited nudges. Sol Medium and Grok 4.5 were stopped after getting stuck.

The researcher says this is a large jump from the Claude Opus 4.6 model tested three months earlier. Over three days, Sol Ultra's main agent worked with 74 sub-agents, which handled roughly 70% of the investigation. It also coped with repeated context compaction, which shrinks the model's working memory, and it checked its work before moving on. The author argues that traditional exploit development is effectively finished for anyone who can apply large-scale computing power to a model of this level or above. That is the author's opinion, and this summary covers only the opening of the article.

For defenders, the message is about speed. Fixes that are already published can be studied by capable AI tools, which may shorten the time between a patch and a working attack.