Eight states sue OpenAI; Anthropic warns of AI worms on home devices

Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each filed suit against OpenAI, alleging the company broke state consumer protection laws by marketing ChatGPT as safe for young people while it encouraged delusions, gave dangerous medical advice and collected data from minors. The complaints cite the wrongful-death suits already filed by families, and even point to this year's rogue AI hacking incidents as evidence of what Arizona calls a "profit-at-all-costs mindset." The states want civil penalties, an end to the safety claims, and in some cases deletion of children's data. OpenAI says it has built "industry-leading age-specific protections," including parental controls and break reminders, and wants rules that apply to the whole industry.

Anthropic's Frontier Red Team says small open-weight models may be months away from being able to copy themselves across cheap, poorly secured internet-connected gadgets. The researchers ran Google's Gemma 4 model split across 15 off-brand Android TV boxes costing about $55 each, and a 9 billion-parameter Qwen model got inference running on eight of nine hardware types it had never seen. Small models still fumble the full chain of finding, infecting and restarting on a new device, but GLM-5.2 and stronger models completed it in simplified tests. Unlike older malware, Anthropic says, such a worm would carry its own weights, adapt to unfamiliar hardware, and have no central switch to shut it off.

A month after publishing logs of the campaigns, Anthropic says its defenses against distillation, where rivals harvest Claude's answers to train cheaper copies, have reduced the data those campaigns collect by 90 to 99 percent. The changes stop newer API accounts from editing Claude's earlier reasoning, limit reasoning transcripts to the organization that requested them, and add a two-stage classifier the company says is as robust as its bioweapons safeguards, with a false-positive rate under 0.1 percent. Anthropic says Zhipu tried to distill Claude Fable 5 while building GLM-5.3, then switched models when the cyber safeguards got in its way, and that it has turned away trillions of tokens of paying traffic as a result.

OpenAI has turned Apple's trade-secret case around, alleging in a Wednesday court filing that a former employee sent more than 100 confidential files to a personal account and then presented them to Apple leadership, including its vice president of operations in Asia. The files allegedly included images of at least one unreleased device, supplier information and specifications for coatings and finishes. OpenAI goes further and links the presentation to the HomeView smart display Apple unveiled this week, though Apple's display had been rumored for years. Apple, which sued OpenAI over hardware secrets in July, did not respond.

Microsoft researchers, with MIT and the European Centre for Medium-Range Weather Forecasts, published a method in Nature that uses machine learning to correct the systematic errors in existing forecast models rather than replacing them. Called Probabilistic Bias Correction, it improved the European Centre's physics model on 98 percent of precipitation targets and 92 percent of temperature targets, and doubled the accuracy of its newer AI forecasting system in the study. It runs on ordinary processors and is open source, and the nonprofit Rhiza Research is testing it with Kenya's national weather service ahead of a strong El Niño rainy season.

Researchers led by Nobel laureate David Baker at the University of Washington used AI design tools to build proteases, the enzymes that cut proteins, from scratch around a known catalytic core, in two papers published in Nature. Their best designs sped up the cutting of a peptide bond roughly a billion-fold while still hitting only their intended targets, and one version cut proteins linked to Alzheimer's disease and frontotemporal dementia inside cells. Therapies are a long way off, since designed proteins are hard to deliver and can trigger immune responses, but the work shows AI design can now produce enzymes aimed at a protein researchers choose.

YouTube Music began showing labels on tracks made with generative AI, sorting songs as Fully Gen AI, Partly Gen AI or No Gen AI using a mix of declarations from music partners and its own detection tools, with the labels also appearing in the main YouTube app. Deezer said it will add a setting before the end of the year that removes AI-generated tracks, albums and artists from search entirely. The two moves land as streaming services struggle with a flood of AI uploads.

Liquid AI released the largest and final model in its LFM2.5 family, a reasoning model for on-device agents that activates just 2 billion of its 24 billion parameters per token, with a context window expanded to 131,000 tokens. Liquid says it is competitive with a larger Qwen 3.6 model from Alibaba on five of eight agent and tool-use benchmarks and runs at 184 tokens per second on an Apple M5 Max laptop. The weights are on Hugging Face under Liquid's own license, which allows free commercial use only for companies with under 10 million dollars in annual revenue, so larger businesses need a separate agreement.