AI News: October 8, 2026

1. Common Sense Media Rated ChatGPT for Teens an “Unacceptable Risk”

Common Sense Media. The organization’s Youth AI Safety Institute ran more than 4,000 test prompts through over a dozen parent-linked teen accounts and found that explicit conversations about suicide, self-harm, and eating disorders never triggered a parental notification. More than one in four situations that warranted a crisis referral did not point users to professional help, and accounts registered as 13-year-olds never switched into teen mode even after several days. The institute wants minors kept off the service until safety is independently verified; OpenAI disputed the methodology, which matters because parental alerts are central to its defense in several teen-harm lawsuits. Source

2. Finland Ordered Google’s Data Center Subsidiary to Halt Site Work

Finnish Licensing and Supervision Agency (LVV). The regulator told Google-owned Tuike Finland to stop land-altering work (tree removal, excavation, blasting, site roads) at its planned Muhos and Kajaani data center sites by October 23, after roughly 330 and 200 hectares of forest were cleared without completed environmental impact assessments. Tuike owes a written explanation by October 14, and officials expect the projects to slip by months. Google said it had “fallen short of our own high standards,” a notable setback for Nordic AI capacity buildouts that have relied on fast permitting. Source

3. Biohub Is Coordinating a $1.8 Billion Effort to Build AI Models of Cell Behavior

Biohub. The Chan Zuckerberg-backed nonprofit’s five-year Virtual Biology Initiative combines Biohub’s $500 million pledge, $300 million from Meta, Google DeepMind, and Isomorphic Labs, and more than $500 million from the US Department of Energy for lab measurements and compute. The NIH is coordinating datasets built with over $500 million in prior federal funding, which Biohub will standardize for model training; the first dataset is due in about a year. Commercial funders get one year of exclusive access to data they paid for before it goes public. Source

4. Nous Research Raised $90 Million at a $1.5 Billion Valuation

Nous Research. The developer of the open-source Hermes Agent closed a $90 million Series B led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital participating, bringing total funding to $158 million. It also launched Hermes for Businesses, which lets companies deploy customized multi-step agents while keeping data private. Nous says Hermes Agent has been cloned more than 24 million times and drives about 2.5 percent of global AI token usage, with annualized revenue of roughly $36 million as of mid-September. Source

5. North American Startup Funding Fell 35 Percent in Q3, but AI Took Two-Thirds

Crunchbase. US and Canadian startups raised $92 billion in Q3 2026, down 35 percent quarter over quarter (mostly because OpenAI and Anthropic had no new megarounds) but up 50 percent year over year, with about $61 billion going to AI-focused companies. The largest rounds were Databricks ($5 billion), Crusoe ($3.9 billion), and Cognition ($2 billion), and there were 11 North American startup acquisitions of $1 billion or more. Seventeen venture-backed companies went public, raising just under $4 billion, with no blockbuster tech IPOs yet. Source

6. AI Startups Have Bought 195 Other AI Companies This Year

Crunchbase. Venture-backed AI companies made 195 acquisitions of AI startups through September 29, already 14 percent above all of 2025, while the number of distinct buyers grew only 2 percent. OpenAI leads with 10 deals, followed by Anthropic and Legora (five each), Harvey (four), and Sierra and Cursor (three each). Only 12 of the 195 deals had disclosed prices, the largest being Nscale’s reported $1.65 billion purchase of Anyscale. Source

7. Mecka Raised $60 Million to Sell Human Motion Data to Robot Makers

Mecka. The two-year-old startup closed a $60 million Series B led by Sequoia, with Nvidia and Microsoft’s M12 participating. Mecka pays people to record everyday tasks such as making coffee or fixing cars while wearing body sensors, then sells the labeled motion data to humanoid robot developers. It is positioning itself as a Scale AI for robotics, competing with XDOF and with LLM data vendors like Scale and Micro1 that are moving into robot data. Source

8. Healthleap Raised $38 Million for Inpatient Risk Flagging

Healthleap. The platform connects to hospital EHRs, uses language models to extract clinical signs from notes, combines them with labs and vitals, and delivers nightly risk scores for conditions like malnutrition and delirium to care teams each morning. It is deployed at more than 50 hospitals, including Penn Medicine, Cedars-Sinai, and Houston Methodist, and its malnutrition program was credited with $23.8 million in annualized impact at the Hospital of the University of Pennsylvania. The funding combines an $8 million seed (Sequoia, First Round) and a $30 million Series A led by Hummingbird Ventures; the company says it flags patients for review rather than diagnosing. Source

9. Automation Anywhere Is Acquiring Boost.ai

Automation Anywhere. The RPA vendor agreed to buy conversational AI company Boost.ai from Nordic Capital for an undisclosed price, with closing expected in Q4 2026. Boost.ai supports more than 36 languages, is strongest in European financial services, telecom, and insurance, and claims 90-plus percent resolution rates in production. The deal follows Automation Anywhere’s late-2025 Aisera acquisition and pairs front-office conversational agents with its back-office automation. Source

10. Nikkei: Chinese Labs Shipped 16 AI Models in September

Nikkei Asia. DeepSeek, Xiaomi, and other Chinese developers released 16 AI models last month, according to Nikkei, despite calls from Anthropic CEO Dario Amodei to slow frontier development as risks mount. The report frames Beijing as showing no sign of slowing, with release cycles among Chinese labs shortening sharply since 2023. Source

11. Researchers Reconstructed Document Pages From ColPali-Style Vector Indexes

Aalto University. A conditional flow-matching attacker reconstructed page images from stored multi-vector indexes with no side information, recovering 47.4 percent of words and 45 percent of sensitive tokens, and re-identifying the source page top-1 among 19,252 pages 98.4 percent of the time. Shuffling vectors did not help (a learned position model restored 93.5 percent re-identification), while pooling at factors of 3 and 9 cut word recall to about 8 percent. For teams running visual RAG, the takeaway is that a raw multi-vector index should be protected like the documents it encodes. Source

12. RoboQuest: The Best Frontier Model Solved Only 23 Percent of Robot Exploration Tasks

RoboQuest. The new benchmark has ten mobile-manipulation tasks in simulated kitchens that require robots to search, inspect, and test objects to find hidden information before acting. GPT-6 Astra led at 23.2 percent success, ahead of Claude Opus 5.5 (13.8 percent), GPT-6.1 Sol (12.2 percent), Claude Fable 5.1 (11.4 percent), and Gemini 3.8 Flash (2.0 percent), even though the models succeeded at isolated execution skills 72 to 81 percent of the time. Missing evidence (stopping exploration too early) caused 43 to 46 percent of failures; the authors released 366 hours of demonstration data. Source

13. openTPU Is an Open-Source LLM Accelerator Designed by AI Agents

openTPU. The Apache 2.0 project, one of the most-discussed items on Hacker News this week, includes SystemVerilog RTL, a custom ISA with a Python simulator, a kernel compiler, and host tools, all produced largely by AI agents. It runs on a Kintex-7 FPGA PCIe card at 133 MHz and decodes Qwen3-0.6B at 31.3 tokens per second (4-bit) and even a streamed-expert Qwen3.5-35B-A3B at 3.95 tokens per second, with device output matching the simulator bit for bit. It is a learning project rather than a production chip, but it is a concrete test of how far agents can go in hardware design. Source

14. A Community Team Formally Verified the Optimal Packing of 11 Squares in Lean

11SquaresFormalized. An AI-assisted project published a Lean formalization proving the optimal packing of 11 unit squares, with a verification run that accepted 7,920 modules and zero admissions. Some numerical checks rely on native_decide, so the result trusts Lean’s native compiler as well as its kernel. It adds to a busy week for machine-checked mathematics alongside large AI-generated proof dumps from frontier labs. Source

15. Artcraft Released Open-Source Clones of Seven Adobe Apps Built With Claude

Artcraft. Developer Brandon Thomas pivoted the AI art tool into a suite of seven open-source Rust apps (with WebAssembly builds) that mimic Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro, describing them as “clean-room replacements” generated with Claude Opus 5.5. The apps are a “super early alpha,” and Hacker News commenters catalogued many gaps. The project is an early test of AI-driven reimplementation of large closed-source commercial software. Source

16. First American Convicted of AI Streaming Fraud Got 18 Months

US Department of Justice. North Carolina musician Michael Smith was sentenced to 18 months in prison after pleading guilty to using AI-generated songs and thousands of bot accounts to collect millions of dollars in streaming royalties. He ran the scheme for seven years before being charged in 2024, and DOJ says he is the first American criminally charged with AI-assisted streaming fraud. Source

17. Melius Raised $20 Million to Generate Ad Creative

Melius. Founded by former Ramp engineers, the startup raised $20 million after scrapping its first product, a tool for managing and optimizing ad spend. It now focuses on generating the creative assets and campaigns themselves, betting that generation, not budget optimization, is where marketers will pay for AI. Source

18. Liquid AI Released Open d1 Decision Models for Edge Devices

Liquid AI. Liquid AI published two open-weight “decision models” on the Hub that return a structured decision in a single forward pass instead of generating tokens: d1-3B (text and image, built on LFM2.5-VL-3B) and the experimental d1-omni-600M (text plus image or audio, built on LFM2.5-Encoder-350M). d1-3B posts an 82.9 mean across seven public datasets versus 81.1 for Decider 4B, and d1-omni-600M scores 78.4 against 77.1 for Decider 2B at about a quarter of the size. Single-question latency for d1-3B is 16 ms on Jetson AGX Thor, 50 ms on Jetson Orin Nano and 8 ms on an RTX 4090; it requires transformers 5.14+ with trust_remote_code, and demos run in the “System One Arcade” Space. Source