China drafts AI blacklists as US Congress spends most of its AI budget on OpenAI

Beijing is moving ahead of Washington. Bloomberg reports Chinese authorities are drafting sanctions against U.S. firms that develop large language models, arguing the tools could be used for cyber attacks on Chinese infrastructure. The specific allegation is that Anthropic blocks access to local businesses but does not guarantee the technology will not be turned against them. The move follows U.S. export restrictions on chips, which China says OpenAI and others are pressing to tighten.

A House report on house.gov shows where legislators’ AI funds went in the latest fiscal year. Congress spent roughly $113,700 on AI services, about 90% of which ($100,600) was paid to OpenAI in 798 separate transactions. Anthropic’s cloud received roughly $13,200. The money funded memo writing, analysis of long bills, citizen-inquiry responses and social-media management. The figures exclude free employee accounts and AI capabilities bundled in enterprise software, so actual usage is broader.

French startup Mistral released Shieldstral, a multimodal content-moderation model with 3 billion parameters under an Apache 2.0 license. Input consists of a prompt system with rules, a violation query, and the material to examine (text, images or a mix). Output is a calibrated probability and a yes/no decision. The company claims safety testing puts it on par with large open models by a factor of seven. A GPU with 16 GB VRAM is enough for local inference, and the weights are hosted on Hugging Face.

OpenRouter’s aggregation platform launched Ori Eval, a tool that scans repositories, identifies model calls and runs candidate models on project-specific tasks such as quality, speed or API cost. The system can evaluate agents, verify function-call correctness, block prohibited actions and generate automatic test cases from free-text bug descriptions. Its goal is to replace public benchmarks with ones that match the actual pipeline for search, code and customer-service workloads.

The Politzer committee reported 8 candidates who disclosed using LLMs this year, the first disclosures since a mandatory reporting rule was introduced. According to Nyman Lev, Wall Street Journal, AP and New York Times are employing LLMs in the research phase for large-scale data processing. Final writing and editing by a model remains prohibited for the candidates.