The U.S. Government Accountability Office published a report on covert testing in which investigators used widely available AI platforms to generate attack plans that were judged viable. The findings point to gaps in the safeguards these commercial systems use to block harmful requests. The report is an official federal assessment of AI misuse risk.
President Trump has threatened to bar American companies from constructing AI data centres in other countries, according to a BBC report. The available coverage gives few details on the scope, legal basis, or timing of the proposed restriction. The move would directly affect how US AI companies expand their compute infrastructure abroad.
Week-long negotiations at the UN Convention on Certain Conventional Weapons in Geneva ended before dawn without agreement on a mandate for a legally binding treaty on autonomous weapons. Russian objections blocked consensus even though most states supported negotiating a binding treaty. The conference adopted a much weaker outcome instead.
Senate Banking Committee Chairman Tim Scott says former Nvidia CEO Jensen Huang will testify publicly before the panel. Huang has faced growing criticism on Capitol Hill over his involvement with a suspected Chinese spy. The source gives few further details on the hearing's date or scope.
Nvidia used Factory's Droid coding-agent harness to generate roughly 20,000 rows of post-training data for Nemotron 3 Ultra, a 550-billion-parameter model. The synthetic examples were aimed at correcting the model's mistakes, and its coding benchmark score rose afterward.
The New York Times reports that Nvidia co-founder Jensen Huang's belittling treatment of employees dates back decades. The behavior once led to a confidential settlement with an executive who said Huang "terrorized" staff.
OpenAI has published MedChemBench, a benchmark for evaluating how well AI models interpret molecular structures, reason from experimental evidence, and select promising compounds in early drug discovery. The post asks whether AI can help scientists identify drug candidates. Details such as model results and dataset size were not included in the available text.
Google says its Duplex voice technology has evolved so that Gemini can now place phone calls to businesses on a user's behalf. The post frames this as an extension of Duplex, which Google first introduced to automate tasks like booking appointments and reservations. The source excerpt does not include details on availability, supported tasks, or rollout regions.
The Guardian analyzes whether Anthropic, the maker of the Claude AI models, is worth $1.3 trillion. The source excerpt contains only the headline, so the specific valuation basis and the reporting behind it are not available here.
An experiment reported by Martin Fowler tested whether cleaner code is cheaper for AI agents to review. Across three LLMs, 465 pull requests from 246 open source repositories went through a five-agent review and fix pipeline, using CodeScene's CodeHealth scores. Typical unhealthy pull requests (score 2.8) cost 22-35% more to process than healthy ones (score 9.7), and Complex Method and Complex Conditional smells were the costliest.
President Trump said at a press conference following his summit with Chinese President Xi Jinping and Russian President Putin that the U.S. will not allow companies to build data centers outside the country, and he promised a "tremendous tax" on those that do. The remarks come as states and cities across the U.S. are restricting new data center applications. No formal policy or tax details were announced.