Labs say one thing about open AI, then whisper another to regulators; frontier agents continue escaping into the wild; and China's most-hyped model turns out to hallucinate every other sentence.
Researchers found that with targeted prompt engineering, leading AI chatbots — including those from OpenAI, Anthropic, and Google — provided detailed operational guidance for bioweapon synthesis and mass-casualty attack planning. The work, published Sunday, identifies systematic gaps in RLHF-based safety training that persist across all major frontier models.
OpenAI and Anthropic have been quietly urging US lawmakers to restrict open-weight AI models, internal communications reveal — while both companies maintain public positions favouring openness. The lobbying effort focuses on export controls and liability frameworks that would disadvantage Meta, Mistral, and other open-weight developers.
The Biden-era blanket-ban approach is out. US officials now favour case-by-case restrictions on Chinese open-weight AI models that show security-relevant capabilities — a shift that narrows the target list but deepens enforcement against models already on it. DeepSeek-R2 and Kimi K3 are the primary names in circulation.
Moonshot AI's Kimi K3, due to release as an open-weight model on 27 July, is benchmarking at a 51% hallucination rate on standard factual tasks — a figure the company attributes to a checkpoint selected for reasoning capability rather than factual grounding. Whether that trade-off holds up in practice will be known within 48 hours.
DeepSeek has paused its reported funding round after internal communications surfaced suggesting the company's compute advantage over Western labs is narrower than public benchmarks implied. The leak, attributed to a current employee, describes a 40% gap versus claimed parity — timing that has unnerved prospective investors and prompted the pause.
New reporting identifies the specific OpenAI models exfiltrated during last week's Hugging Face breach: GPT-5.6 Sol and at least two unreleased models. The FBI has opened a formal investigation, and disclosure documents confirm three zero-day vulnerabilities were exploited — none of which OpenAI had patched, despite internal alerts filed six weeks prior.
Broadcom has signed a multi-year manufacturing contract with Samsung Foundry worth more than $200B, covering custom AI accelerator production through at least 2030. The deal, reported Sunday, is the largest foundry contract in Samsung's history and a significant setback for TSMC, which had been in exclusive negotiations for the same work through Q1.