California AG Rob Bonta sued OpenAI on Monday in San Francisco Superior Court, alleging it ran dangerous AI agents without proper safeguards. Bonta cited the July Hugging Face breach and said OpenAI later disclosed 80 additional 'loss of control' incidents involving third-party networks and websites.
AI news for Tuesday, November 17, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
Five developments: one new today and four from yesterday's alerts.
Pick the topics, companies and people you track. We’ll follow what leading AI experts are reading and sharing, filter the noise, and send your focused edition weekly—or sooner when enough important news breaks (up to three times per week).
Choose my signals →The rest of today’s expert-filtered signal, ranked for consequence, utility, surprise, and range.
NATO's Communications and Information Agency has consumed up to 300 billion OpenAI tokens a month through the Daybreak defender program, head of national security policy Sasha Baker disclosed. ENISA used the models to find four vulnerabilities in EU institutional software; Ukraine, France, Germany and Poland also have access.
A Reuters review of more than 300 Chinese papers, patents and filings found distillation of US frontier model outputs has moved from academic experiments to mainstream industrial use. The records name OpenAI's o1, GPT-4, GPT-4o and GPT-4o mini among the model outputs mined to train Chinese rivals.
Anthropic told IPO investors that Claude Mythos Preview ported nearly all the kernels needed to run one Anthropic model on AMD MI355X GPUs in about three days with 'basically no human intervention', per Tom's Hardware. Chief compute officer Tom Brown said the same work historically took dozens of engineers six to twelve months.
Anthropic's IPO filings show $900.9 billion in infrastructure commitments through 2045, per the FT, including $7.2 billion a year for five years on the first 1 GW tranche of Google TPUs across four sites. The new 'Anthropic Direct' plan lets Anthropic lease its own data centers rather than rely only on AWS, Google and Azure.
Alphabet's Isomorphic Labs closed a $5 billion Series C co-led by Sequoia Capital and Greenoaks, Bloomberg reports, 2.4x its $2.1 billion May round and about 8x its March 2025 $600M round. Andreessen Horowitz joined as a new investor alongside Thrive Capital, Alphabet's GV and CapitalG, Temasek, MGX and the UK Sovereign AI Fund.
A Harvard-led study in Nature Sustainability estimates 403 US hyperscale data centers consumed 68 to 99 terawatt-hours from May 2024 to April 2025 and emitted roughly 25.7 million metric tons of CO2, comparable to Denmark or Bolivia. That is about 2% of US electricity, more than New York City and Los Angeles combined.
Useful releases, methods, and workflows worth trying.
MBZUAI released K2-Horizon, six fully open language models spanning a 0.9B dense network through a 375B-A23B mixture-of-experts flagship. Every intermediate training checkpoint, roughly 10 trillion synthetic pretraining tokens, and the pipelines that produced them ship with the weights.
Muhammed Nazeem published tiny-voca, a 4.5-million-parameter English TTS trained from scratch on a single RTX 4090 in 7.7 hours for $9.56 of GPU time. The model card reports the checkpoint runs about 70 times faster than real time on four CPU threads, with ONNX weights matching PyTorch audio to within 0.6% of peak amplitude.
That’s today’s shot. — Alexis · AI Weekly · Sponsor this newsletter
