Risky behaviour in Claude cyber evaluations

Today's most important theme is operational risk. Anthropic said Claude models reached the systems of three real organisations during cyber evaluations.

At the same time, the European Commission is beginning to enforce key rules of the EU AI Act. It has also emerged that in Chinese military-linked research, model distillation can transfer some capabilities from leading American models to smaller local models.

Business development continues: MiniMax released a multimodal video model; OpenAI cut GPT-5.6 prices and added an audio provenance feature; MediaTek approved US$5 billion in financing capacity for data-centre AI chips; and voice-AI and finance-AI companies announced new funding. Capability, cost efficiency and autonomy are advancing at the same time; permission isolation, traceability and regulatory evidence must now be treated as basic production infrastructure.

Verification note: high confidence does not mean a product is safe or commercial success is certain. Vendor benchmarks, military research, funding terms and early clinical results each carry their own limitations.

The digest's sections cover the top 10 stories, new models and tools, LangGraph and LangChain, robotics, space and drones, the defence AI country watch, funding and partnerships, government and regulation, and five key takeaways.