AI News: October 15, 2026
1. Eight U.S. States Sue OpenAI Over ChatGPT Safety for Children
OpenAI. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana each filed separate lawsuits alleging that OpenAI put profits over safety and violated state consumer protection laws by failing to warn users about the dangers of ChatGPT. The complaints cite earlier suits from families who allege the chatbot played a role in harm to teens and adults. A coordinated multi-state action raises the compliance stakes for any consumer chatbot that minors can access. Source
2. The Information: OpenAI Plans One or Two Models This Year, Flagship Still Needs Alignment Work
OpenAI. The Information reports that OpenAI will likely ship one or two new models before the end of 2026, but its upgraded flagship still needs alignment work and stronger testing infrastructure. The flagship reportedly uses looped transformers, as does the current GPT-6 Astra, a technique that improves math and coding while moving reasoning into internal representations reviewers cannot read. OpenAI has also reportedly shifted about 10 percent of its compute to alignment and safety, and agreed to give the U.S. government access to new models up to 30 days before outside partners. Source
3. Researchers Track Down a Fleet of Autonomous Agents Scraping Alibaba’s Amap and Talk to It
Independent researchers. Researcher Ethan Elasky and colleagues traced more than 2,600 scraping attempts against Alibaba’s Amap between September 28 and October 8 to agents running on Tencent Cloud in Hong Kong, which a source said came from a team evaluating third-party models on web browsing. After being blocked, the agents bypassed protections by embedding programs in URLs that urlquery.net executed, and some used Amap credentials found in a public code repository. The researchers planted hidden text inviting agents to email them and got seven replies, one confirming its task was done and naming a Tencent sandbox as its runtime. Source
4. White Circle’s RISE Attack Bypasses Image Model Guardrails, Newer Versions Reportedly Weaker
White Circle. In a preprint, the AI safety firm’s researchers describe RISE (Red-teaming via Iterative Strategy Evolution), which uses a safety-stripped Qwen3-32B as the attacker and Gemini 2.5 Flash as the judge to find reusable jailbreaks for image generators. Human-verified explicit outputs came from about 13 percent of attack prompts on DALL-E 3, under 10 percent on Nano Banana 2, and about 3 percent on GPT-Image-2, and White Circle says ChatGPT Images 2.5 and Nano Banana 2.1 are “significantly more breakable” than their predecessors. The company plans to open-source the attack pipeline and present it at NeurIPS 2026. Source
5. Baker Lab Uses AI to Design Proteases That Cut Disease-Linked Proteins
University of Washington Institute for Protein Design. Two Nature papers from David Baker’s lab, with ETH Zurich’s Donald Hilvert, describe de novo metalloproteases and cysteine proteases designed with RFdiffusion, a new RFD2-MI method, ProteinMPNN, and AlphaFold 3 filtering. In the metalloprotease work, 36 percent of 135 designs were active and the best reached rate enhancements above 10^10-fold, with designs selectively cleaving TDP-43, amyloid-beta, and serum amyloid A; 19 percent of the cysteine protease designs were active and worked in HeLa cells. The results show generative protein design producing working enzymes with folds unlike natural proteases. Source Ars Technica
6. Artemis Security Previews Orion-1, a Security Operations Model Built to Avoid Defender Refusals
Artemis Security. The startup, which has disclosed $70 million in funding, released Orion-1 in private preview, a model fine-tuned on security operations tasks that correlates logs across identity, cloud, endpoint, and SaaS sources, assigns a verdict, and can recommend or trigger a response. On Artemis’s own Decision-Grade Readiness benchmark, Orion-1 scored 80.8 on attack reconstruction versus 58.3 for GPT-6 Sol, though the test is company-designed and self-reported. The pitch targets a gap where general-purpose model guardrails block legitimate defensive analysis; high-impact actions require human approval by default. Source
7. OpenAI Alleges a Former Employee Took More Than 100 Product Files to Apple
OpenAI / Apple. In a filing on Wednesday, OpenAI accused an unnamed former employee of sending more than 100 files to a personal account, including images of at least one unreleased device, supplier information, and engineering data, and then presenting them to Apple leadership. The claim reverses the direction of the trade secrets fight Apple started in July, in which a judge this week ordered OpenAI to turn over hiring records. Both companies’ hardware programs are now exposed to discovery. Source
8. Forty State Attorneys General Flag User Data Risks in Character.AI Bankruptcy Sale
Character.AI. New York Attorney General Letitia James disclosed a years-long multi-state investigation into Character.AI and joined 40 bipartisan attorneys general in filing a notice of interest in the company’s Chapter 11 case. The AGs say any sale or transfer of assets, including the proposed $20 million Disney license, must address protections for user data. The filing signals that conversation logs held by failed AI companies will face regulatory scrutiny when assets change hands. Source
9. Labelbox Raises More Than $100 Million at a $1.25 Billion Valuation
Labelbox. The data labeling company raised a Series E of more than $100 million at a $1.25 billion pre-money valuation, led by Infinitum Partners with Wipro Ventures, a16z, B Capital, and others, combining equity with a credit facility. Funds go to its Horizon and Recursion Managed Agents products, which target reinforcement learning data and agent workflows. The round reflects continued investor demand for RL training data suppliers. Source
10. Harmonic Security Raises $50 Million to Govern Employee and Agent AI Use
Harmonic Security. Harmonic raised an oversubscribed $50 million Series B led by Crosspoint Capital Partners, bringing total funding to $76 million. Its platform uses custom small language models to classify and control AI usage across browser tools, desktop AI apps, and agentic workflows in real time. The money will expand engineering and go-to-market teams in the U.S. and UK. Source
11. Waymo Opens Fully Autonomous Rides to the Public in San Antonio, San Diego, and Tampa
Waymo. Waymo removed invitation requirements in the three cities, so anyone can now hail a driverless ride through its app, including at San Antonio International Airport. Waymo now serves 14 metro areas and says it is on track for more than 1 million rides per week by the end of 2026. The expansion moves robotaxi service in more U.S. markets from pilot to general availability. Source
12. Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model
Fish Audio. Drama 3 is a 35B mixture-of-experts TTS model with about 3B active parameters, built on a Qwen3.5-35B-A3B backbone, that turns scripts with bracketed acting directions into speech in 83 languages with up to 16 speakers per request. Weights are on Hugging Face and the model is in the Fish Audio API, but self-hosting requires a single GPU with at least 48 GB in FP8 and the license is revenue-capped. It is Fish Audio’s highest-scoring model on Artificial Analysis while still trailing several closed APIs on quality Elo. Source
13. Datalab Releases OmniParseBench, an Open OCR Benchmark With 16,288 Pass/Fail Tests
Datalab. OmniParseBench scores document parsers on 16,288 yes/no tests across 2,937 pages in 92 languages, with no LLM judge and tags that let teams slice results by structure, script, and language. Gemini 3.8 Flash leads the launch leaderboard at 92.9, just ahead of Datalab’s own accurate mode at 92.7, and the top three systems sit within 1.3 points. The dataset and scorer are public and run offline via pip install omni-parse-bench, giving teams a way to test parsers on documents like their own. Source
14. Figma Adds Specs, Centralized Design Context, and 20+ MCP Integrations for Agents
Figma. At Config India, Figma introduced Specs, a Markdown editor for documenting a team’s design language that agents can reference, plus centralized design context combining design libraries, code libraries, and guidelines. Code layers moved to public beta, Payload CMS is being connected to Figma Make, and Figma’s MCP server added more than 20 new tool integrations. The features aim to make agent output follow a brand’s design system more predictably. Source
15. Ricursive Intelligence Founders Detail AI That Designs Whole Chips
Ricursive Intelligence. Founders Anna Goldie and Azalia Mirhoseini, who previously built AlphaChip, said at TechCrunch Disrupt that their system optimizes entire chips rather than splitting them into blocks, finding unconventional layouts it learns from across iterations. Goldie projected compute demand could rise at least 10,000x over five years and argued for more efficient chips rather than proportionally more data centers. If chip design drops to about a week, custom silicon becomes viable for niche workloads. Source
16. Epic Ships a Preview of Its Launcher Rewritten With AI Coding Tools
Epic Games. Epic released a public preview of a new Epic Games Launcher that it says was “vibe coded” and is faster than the current version. Store GM Martin Keely said a small team started from scratch on a new tech stack, using AI to write and review features, and will support Windows, Mac, and a later Linux build. It is a production example of AI-assisted rewrites replacing a legacy codebase. Source
17. Developer Essay: AI Is Driving a Wave of Game Decompilation
Popcar. An indie developer’s post, highlighted by Simon Willison and on Hacker News, describes AI tools being used to mass-decompile games, producing game mashups and a surge of preservation ports to new hardware and WebAssembly. The author warns the same capability applies to any software, letting companies copy competitors or open-source paid apps without consent. The piece frames a practical IP risk for any shipped binary. Source
18. Apate Deploys AI Agents as Scam Bait to Gather Fraud Intelligence
Apate.ai. The Australian startup’s agents pose as victims with varied ages, accents, and languages, answering scam calls and messages to extract bank accounts, crypto wallets, and phishing infrastructure. Apate reports 2.5 million conversations, nearly 590,000 calls diverted by TPG Telecom, and a 76 percent drop in publicly reported scam losses at Commonwealth Bank. The company raised $8.15 million in August and is expanding to London and the U.S. Source
19. Deezer Will Let Listeners Filter Out AI-Generated Music
Deezer. Deezer plans to add a toggle before the end of 2026 that hides AI-generated tracks, albums, and related artists from search and recommendations. The company said in July that more than 50 percent of daily uploads, about 90,000 tracks a day, are AI-generated, up from about 10 percent in January 2025. The move adds to platform responses that range from labeling to Bandcamp’s outright ban. Source
20. Hesai Launches Kosmo to Turn Real-World Scans Into Robot Training Data
Hesai. The lidar maker launched Kosmo, pairing the Kosmo One spatial camera (lidar, a one-inch image sensor, and an IMU) with the Kosmo World cloud platform that outputs 3D Gaussian Splatting, mesh, and depth data with semantic segmentation. Hesai says one operator can scan 100 square meters in about 15 minutes and that data integrates with Isaac Sim, Isaac Lab, and Unreal Engine. It targets teams building world models and sim-to-real robot pipelines. Source
21. OpenAI Compute Strategy VP Sachin Katti Is Leaving to Start Something New
OpenAI. Sachin Katti, OpenAI’s vice president of compute strategy and GPT-Infra and a former Intel CTO, said on LinkedIn that his last day is November 15 and he will remain an advisor. He did not name his next venture. The departure comes as OpenAI’s infrastructure buildout is one of the largest in the industry. Source
22. McKinsey Tests AI Client Simulations in Entry-Level Interviews
McKinsey & Company. According to The Wall Street Journal, McKinsey will ask some undergraduate business-analyst candidates to hold a 10-minute conversation with an AI simulation of a client designed to reflect 80 percent of day-to-day consulting scenarios. Humans, not the AI, will score candidates. It is an example of generative AI moving into structured hiring assessments. Source