Chaining browser bugs into a sandbox escape was work for a few specialist teams. Anthropic says its model did it with little human involvement and also built working attacks on Safari, Android, Windows and Linux, software used by billions. Read: Chaining browser bugs into a sandbox escape was work for a few specialist teams. Anthropic says its model did it with little human involvement and also built working attacks on Safari, Android, Windows and Linux, software used by billions. Read: Epoch says the model built working exploits for six Firefox versions from fixes public in nightly builds but not yet released, at about $1,000 each. That attack once took skilled people or restricted frontier models, so the patch gap is now cheap to use. Read: Artificial Analysis scores it 26, one point above Muse Spark Open, making it the strongest open model built outside China. It still trails DeepSeek V4.1 Flash and GLM-5.3 by 6 points, and selling it as a hosted service needs a separate Mistral license. Read: We covered the escape itself on Oct 24. OpenAI now says probes, reviewer models and investigative agents watch all frontier RL runs and internal Codex use, and a replay of its earlier Hugging Face breach alerted 30 hours before the agents got in. Read: Pipecat stays BSD licensed under a new independent foundation, so teams building on it keep their choice of models and vendors. Deepgram also proposed an open protocol that carries tone and speaker cues between voice models instead of only the words. Read: Image-heavy requests such as long computer-use sessions and large PDFs were hitting the 32 MB cap with context window still unused. Four times the room lets those sessions run longer before the request size, not the model, ends them.
Unreleased Claude model turned a Linux kernel bug into a remote root exploit