Anthropic will watermark text from future Claude models

Anthropic says future Claude models will add a statistical watermark to generated text worldwide. The method changes low-stakes word choices without adding hidden characters, extra tokens, or identifying data. Detection estimates whether Claude helped write a passage; short, edited, factual, and exact-code samples carry weaker signals. 1

Z.ai released GLM-5.3 for hosted coding access, reporting a Terminal-Bench 3.0 increase from 4.6 to 28.3 and an ExploitBench increase from 24.4% to 54.4%. The company says open weights will follow after roughly two weeks of safety evaluation and hardening. 2

Google showcased HEIR, an open-source compiler that converts pre-trained models to run on homomorphically encrypted data. Its CPU demonstrations covered recommendations, fraud detection, network anomaly detection, and hotword recognition; compute cost, latency, and model compatibility remain the practical constraints. 3

Reuters reports that Apple trained a China-specific language model with Alibaba's support. Alibaba's Qwen is still expected inside Apple Intelligence on compatible devices in China, while the role and release timing of Apple's own model remain unannounced. 4