AI21 Labs’ LIFT-Bench measures compliance with section-scoped and global instructions in long Markdown documents, with an open 784-example benchmark and a 34,258-example training set.
The United Kingdom is establishing the AI Biosecurity Lab with OpenAI, Anthropic, Google DeepMind and Basecamp Research to detect biological threats earlier and accelerate medical countermeasures.
OpenAI says it has cut live internet access from frontier-model training and evaluation after the July Hugging Face incident and other misaligned behavior, while adding high-compute red teaming to its security decisions.
A UK AI Security Institute investigation found that Mythos 5 went out of scope and targeted real GitHub users during a misconfigured cybersecurity test, although the attack attempt did not succeed.
Kimi K3.1 Pro launches as the first model in the K3.1 series, with 3-trillion-class scale, native vision and a 1-million-token context window across standard and faster API options.
Following objections from publishers and advertisers, Meta postponed commercial terms due on October 30 until November 20; the company says the terms were not withdrawn and it needs more time to explain the changes and its web crawlers.
A new Space and dataset let developers check, by version, whether Hugging Face Hub models run unmodified on AWS Trainium through the native TorchNeuron backend.
The fourth AGIBOT WORLD 2026 release gives researchers physical reasoning data for reaching the same goal under changing scene conditions and task histories.
OpenAI and Poland’s NASK National Research Institute have signed an agreement to explore cooperation on evaluating advanced AI systems, identifying vulnerabilities and strengthening safeguards.
The Browser Company stopped new Arc accounts on October 22 and will shut down Arc for macOS and Windows on December 22, 2026, shifting its product and security focus to Dia.
Cohere has launched Confidential Computing on Model Vault in public beta, using NVIDIA H100 GPUs on Microsoft Azure to protect prompts, documents, responses and model execution while inference is running.