ROK-FORTRESS benchmark tests AI safety across languages

Built by a Reddit user who found ComfyUI node-wiring slow and models too large for their GPU, EasyComfee bundles Image, Video, VFX, Lip Sync, Music, Voice and Genjutsu modes into one studio.

Since GPT-5.6, OpenAI has run activation classifiers on all assistant completions, reading a model's internal activations to flag misaligned behavior. OpenAI says probes block policy-violating turns mid-stream and resist jailbreaks better than much larger text classifiers.

Apple is developing an AI-driven roaming home robot, code-named J595 and possibly named HomeView Ultra, targeting a $1,000+ price by 2027, with delays possibly pushing it to 2028. Prototypes date to 2018; features reportedly include FaceTime calls and automatic lighting.

OpenAI CEO Sam Altman spoke at the ADIPEC opening ceremony in Abu Dhabi, covering AI's future, energy innovation and human control of the technology. He highlighted the UAE's growing role in AI adoption.

LifeVal measures how well AI helps people complete daily tasks like finding childcare, managing finances, and handling government paperwork. OpenAI says the benchmark was shaped by feedback from thousands of ChatGPT users on where they most want assistant help.

A rule named "tengu_heron_brook" in Claude Code 2.1.314 (released Oct 30) tells Claude to build original or genre games but not recreate a specific existing game, even if renamed. Anthropic engineer Addy Osmani confirmed two sentences were added to Opus 5.5's prompt because using Claude to infringe IP violates usage policy.

NVIDIA's AI Cluster Runtime, part of the DSX AI Factory Platform, captures tuned configs as reusable recipes and validates running clusters so Kubernetes is ready for Vera Rubin NVL72 compute trays at launch. Vera Rubin requires a new GPU driver branch, validated with host-installed R615.23, with early support targeted for R610.