AI Daily · 2026-08-28

China's frontier labs rolled out competing flagship models: Zhipu's GLM-5.3 credits post-training for all gains and claims open-weight SOTA on coding and vulnerability benchmarks, while Tencent's Hy4 preview pairs a 770B MoE with 1M context and narrowly led GLM 5.3 and Kimi K3 in an internal engineering-task blind test. Both releases emphasize complex coding, long-horizon tasks, and agentic or productivity workflows, shifting attention beyond generic chat. In security news, a demonstrated prompt-injection attack against Claude Code「Auto Mode」reportedly succeeded about 80% of the time, underscoring the need to sandbox unattended agents and keep credentials out of their environment. Also today, Hugging Face added Hindi and Indian English to its Open ASR Leaderboard, Anthropic opened Claude for Teachers to U.S. K-12 districts, and NVIDIA announced an expanded GPU infrastructure partnership with AWS.

North America · First-hand

Anthropic

⭐⭐ [Product Update] Claude for Teachers, now available for schools and districts

Anthropic is expanding Claude for Teachers into a free Enterprise offering for U.S. K-12 schools and districts, letting administrators manage educators and staff in one centrally organized account with features like single sign-on, role-based access controls, and domain claiming. Schools and districts must verify and accept K-12 terms and a data privacy agreement; Claude for Teachers data is not used for model training and is covered by FERPA-aligned commitments. For the new school year, Anthropic added two teaching skills co-developed with Learning Commons - lesson preparation and check for understanding (math at launch) - plus improvements to existing skills and an updated Claude for K-12 Academy with courses, guides, and classroom workflows. This fall, the company is piloting an evaluation in the Detroit Public Schools Community District to study effects on educator wellbeing and practice, and qualifying organizations can sign up by June 30, 2027 for a year of free access.

East Asia · First-hand

Zhipu GLM

⭐⭐⭐⭐ [Model Release] zai-org/GLM-5.3

Zhipu AI released GLM-5.3 on HuggingFace, a 753B-parameter text-generation model built on the same base as GLM-5.2, with all gains coming from post-training. The vendor claims it is the most capable open-weights model for coding, showing a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench and achieving open-source SOTA on public benchmarks such as Terminal Bench 3.0 and Agents' Last Exam. Post-training also led to emergent cyber capability, with SOTA on CyberGym for vulnerability discovery and more than double GLM-5.2's score on exploitation benchmarks. The model can be deployed locally with SGLang, vLLM, Transformers, and other frameworks, and supports a reasoning_effort parameter to control thinking budget.

⭐⭐⭐⭐ [Model Release] zai-org/GLM-5.3-BF16

Zhipu released GLM-5.3-BF16, an open-weights text-generation model with 753B parameters. It shares the same base model as GLM-5.2, with all gains coming from post-training, bringing much stronger complex coding and long-horizon task performance: a 50% improvement over GLM-5.2 on Z.ai Code Bench, plus open-weights SOTA on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam. The model also shows emergent cyber capability, reaching SOTA on CyberGym for vulnerability discovery and more than doubling GLM-5.2 on exploitation benchmarks. GLM-5.3 supports local deployment via frameworks such as SGLang and vLLM, and exposes a reasoning_effort parameter (low/high/max, default max) to control the thinking budget.

Tencent Hunyuan

⭐⭐⭐⭐ [Model Release] tencent/Hy4-preview-FP8

Tencent's Hunyuan team released Hy4 preview, a new flagship Mixture-of-Experts model, along with an open-source FP8 quantized version under the Apache-2.0 license. The model has 770B total parameters with 49B activated per token and supports a 1M context length, featuring Gated DSA attention, iHC residual connections, and a native MTP layer for speculative decoding. It was co-designed with Tencent product teams and optimized for software engineering, office/analysis, game development, and scientific research workflows. In an internal blind evaluation, 163 experts rated it slightly ahead of GLM 5.3 and Kimi K3 on 203 engineering tasks (2.99 vs 2.92 and 2.94 respectively). Tencent notes this is an early version with remaining headroom in both pre-training and post-training.

⭐⭐⭐⭐ [Model Release] tencent/Hy4-preview

Tencent released Hy4 preview, a new-generation MoE flagship model with 770B total parameters, 49B activated per token, and 1M context length. The architecture features Gated DeepSeek Sparse Attention with IndexCache and iHC residual connections, plus a native MTP layer for speculative decoding. The model is scaled up in model size, context length, and training data, with a focus on productivity scenarios such as software engineering, office/analysis, game development, and scientific research. In an internal blind side-by-side evaluation across 203 engineering tasks, Hy4 preview scored 2.99, slightly ahead of GLM 5.3 (2.92) and Kimi K3 (2.94). Tencent notes this is an early version with headroom in both pre-training and post-training.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Research

⭐⭐⭐ [Research] The Open ASR Leaderboard Adds Its First Global South Language

Hugging Face and Voice Arena have added two evaluation sets to the Open ASR Leaderboard: Monsoon en-IN and Monsoon hi-IN, targeting Indian English and Hindi. Hindi is the first Global South language on the multilingual tab, which previously covered only European languages. Each set has a public and a private split; the four splits are speaker-disjoint and cover 4,888 speakers with 12 speaker attributes each. The dataset was designed to vary along nine axes — geography, age, gender, vocabulary, devices, acoustic environments, speech type, speech rate, and multiple valid transcripts — to expose disparities hidden by aggregate WER. Collection involved recruiting across hundreds of districts, using contributors' own handsets and connections, and prompts covering everyday topics.

⭐⭐ [Research] Breaking Claude Code Opus 5 Auto Mode

Simon Willison's link blog reports on Johann Rehberger's attack against Claude Code's auto mode. Anthropic recently made auto mode the default and made strong claims about its protection against prompt injection. Rehberger's attack, claimed to work about 80% of the time, tricks Claude Code into downloading and unzipping an archive, then importing base64 in a way that actually executes a local struct.py file extracted from the archive. In some runs, auto mode blocked Claude's own cleanup commands even after the model detected the compromise. Simon agrees with the conclusion that sandboxing is the only safe option when adversarial attacks are a risk: run unattended agents in containers, VMs or OS sandboxes, restrict network egress, monitor agents, and avoid exposing home directories, SSH keys and cloud credentials.

Other

⭐⭐ [Other] NVIDIA Announces Upcoming Event for Financial Community

NVIDIA announced a collaboration with AWS to deliver 2 million additional GPUs and next-generation infrastructure for agentic and physical AI. The announcement is dated August 26, 2026.