AI News from September 12, 2026

Anthropic’s CEO reportedly warned rogue AI agents could disrupt the internet within months and backed a plan to slow deployment of powerful systems.

Anthropic’s CEO is urging AI companies and governments to pace frontier development, arguing that safety work must keep up with faster, more capable models.

Hugging Face’s Sentence Transformers v6.0 adds multi-vector retrieval and training, giving developers a practical path to higher-quality domain search at higher index cost.

A KAIST and Naver AI Lab study finds models encode distinct reasoning operations internally, raising new possibilities and limits for AI oversight.

AWS adds model caching to SageMaker HyperPod and prefix-aware routing to SageMaker Inference, targeting faster scale-out and lower LLM latency.

Y Combinator CEO Garry Tan reportedly told OpenAI and Anthropic to avoid reacting to an alleged Chinese AI model, highlighting restraint in AI rivalry.

Harvey has reached a $15.5 billion valuation in a new funding round, signaling continued investor demand for legal AI despite limited disclosed deal details.

The Institute of Foundation Models has released K2 Horizon models with weights, code and training data, expanding access to inspectable AI development.

Y Combinator CEO Garry Tan backs an American regime allowing open-weight labs to distill frontier models, challenging restrictions and raising policy questions.

OpenAI says GPT-6 Astra brings computer use, coding and enterprise controls to ChatGPT Work, Codex and its API, with launch claims still unverified.