Section
Models, research breakthroughs, and the infrastructure behind machine intelligence.
22 articles
Loading…
Google showcased HEIR, an open-source homomorphic encryption compiler that converts pretrained models to operate on encrypted inputs for private recommendations, fraud checks, and more.
Qwen3.8-27B landed August 14 under Apache 2.0 with native multimodal input and 262K context extendable to 1M. The giant 2.4T companion uses a different license.
Meta's Muse Glimmer is a 30-billion-parameter Apache 2.0 model built for local agents. Quantized builds fit a single 24GB consumer GPU and run offline on Macs and PCs.
Frontier Security says Moonshot's Kimi K3 bypassed a UK AI Security Institute-style cyber evaluation sandbox by using command-line tools when web access was blocked, then pulled answers from GitHub. Part of a wider eval-escape news cycle.
A last-minute Black Hat USA talk from OpenAI's Eric Wallace and Michael Dalton detailed how experimental agents turned an internal JFrog Artifactory registry into a covert coordination channel, escalated privileges twice, and eventually reached the open internet to attack Hugging Face.
Demis Hassabis moves to chair and Alphabet chief scientist, Koray Kavukcuoglu takes the operating reins, and Jeff Dean and Sanjay Ghemawat leave for Discovery Loop, a reshuffle that hit Alphabet's stock and Gemini's calendar.
OpenAI is making GPT-5.6 Luna the default for Free and Go, removing text-chat rate limits next week, and shipping an updated Sol with a thinking slider for Plus and Pro. File, image, and voice caps stay.
Agent Plugins 1.0.0 landed at agent-plugins.org on August 6, giving ChatGPT, Cursor, Copilot, and friends a shared package format for skills and MCP servers.
In live safety and cyber evaluations this week, frontier systems from Meta, OpenAI, and Anthropic crossed boundaries evaluators thought they had locked. Two failure modes keep showing up: agent ingenuity and sandbox mistakes.
