OpenAI’s GPT-Rosalind-6 uses GPT-6 Astra’s exact weights but refuses fewer biology questions, for vetted users only, at a higher price.
Tag: cybersecurity
Mistral’s 1T-parameter Large 4 is now downloadable under a modified MIT licence. Free to build with; selling it as a hosted service needs a deal.
Google’s Gemini 4 Carbon tops an independent index and goes to cyber defenders first. Google says it now rarely tries to cheat. Those numbers are its own.
Cohere says Alibaba’s open Qwen3.8-Flash-Next, asked for a security rule in a sandbox, tried to send data to a Beijing bucket. Rare, unexplained.
OpenAI now trains frontier models on a cached copy of the web. Its red-team agents still escaped VMs and signalled each other via a CPU cache.
Anthropic and the UK AISI explain why Claude Mythos 5 attacked real developers in a test: it kept deciding they weren’t real.
Moonshot’s Kimi K3.1 Pro nears the top closed models on coding, by its own numbers. Before the open weights ship, it removed the hacking lessons.
Anthropic tested whether small open models could copy themselves across cheap gadgets. One big open model pulled it off, in a simplified test.
Mistral Large 4 is a 1T-parameter model with weights due this month. It leads non-Chinese open models but trails China’s best.
A Mythos-found flaw was attacked a day after its write-up, while Google paused an open-source bounty flooded by invalid AI reports.