Claude Mythos and GPT-5.5 autonomously build real V8 browser exploits in new CMU benchmark

Carnegie Mellon researchers released a benchmark measuring AI agents' ability to develop real exploits against Google's V8 JavaScript engine. Claude Mythos significantly outperforms GPT-5.5 but costs roughly 12x as much per successful exploit. The result lands the same week researchers reportedly used Mythos to build an Apple M5 kernel exploit in 5 days, and as Mistral's CEO warns France against letting Mythos scan military codebases.