AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Google's Threat Intelligence Group found an exposed server running an AI framework called "exploit_agent" that discovered, built and ran exploits against vulnerabilities with no human involvement, which GTIG describes as a first.

OpenAI physicist Nathaniel Craig used GPT-6 Astra to turn his strategy for putting the full Standard Model on a lattice into a working proof, addressing a phase-setting problem that has blocked overlap fermions for over 25 years. Craig cautions that realistic simulations remain distant.

Moonshot AI is shifting its Kimi models from coding benchmarks toward biology, chemistry and pharma research, citing diminishing returns in coding. Partner DP Technology is integrating Kimi K3.1 Pro into its BioMaster research assistant.

Anthropic's research post reports that Claude agents retain more across sessions when they save a few strategy notes rather than exhaustive logs, and that a post-session "Dreaming" pass can correct faulty memories. In a Pokémon Red trial, Claude Fable 5 earned four badges versus one for the older Sonnet 3.7, which had written 413 files.

Anthropic's Claude Fable 5.2 creates images by writing Python programs that simulate physical brushstrokes on a virtual canvas rather than computing pixels, producing a photorealistic parking garage scene from roughly 700,000 strokes. It scored 64.9 percent on Anthropic's internal RoboDojo Sim benchmark against 29.8 percent for OpenAI's GPT-6 Astra, though Astra painted faster in a real robot-arm test.

Australia's Reserve Bank says early data shows job postings and employment falling fastest for 15- to 24-year-olds in highly AI-exposed fields such as software development, clerical work and customer service since ChatGPT's late-2022 launch. The author cautions the analysis shows correlation, not causation.

Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter agent model it says was trained entirely on Chinese hardware, with a sample deployment running on 194 Huawei Ascend chips. It scores 88.4 on Terminal Bench, just behind Claude Opus 5.

Anthropic says Claude Opus 5.5, now the default on all paid plans, performs best when users drop legacy phrases like "think hard" and start at the default Medium effort level, which the company says matches its predecessor Opus 5 on High.

Former Anthropic researcher Jacob Coxon says a slide from Anthropic's IPO roadshow shows algorithmic efficiency improving at an accelerating rate and was a key reason he left the company after about four months.

Runway released Praxis-1, an open-source robot control model with weights and code under Apache 2.0, and it leads the World Action Model category on Nvidia's RoboLab-120 benchmark at 46.2 percent success. Runway ran the tests itself and they are not yet independently verified.

Nvidia reportedly built a credit-guarantee program under Jensen Huang that backed AI cloud providers' chip purchases, including talks with OpenAI over a $250 billion guarantee; interim CEO Colette Kress later halved that commitment, and a Wall Street Journal report links Huang's ouster more to these business practices than to his personal affairs.

OpenAI has launched ChatGPT Wallet for US Pro subscribers, letting ChatGPT agents pay for online purchases after each one is approved with Face ID or a fingerprint; card numbers are replaced by tokens from Amex, Mastercard and Visa.

Banks leading Anthropic's IPO, including Goldman Sachs and Morgan Stanley, are rejecting buy orders from mainland China and Hong Kong, with some citing US ITAR export controls, per Bloomberg. The stock is expected to trade on Nasdaq as ANT after pricing on November 18.

Anthropic released Claude for Manufacturing, a plugin that links Claude to Autodesk Fusion, Oracle NetSuite, Microsoft 365 and Infor, adding 17 skills for tasks like drafting quotes, comparing suppliers and arranging backup orders after human approval.

US public-sector unions are securing AI guardrails through collective bargaining and legislation, including retraining instead of layoffs in Oakland and a ChatGPT Enterprise pilot in Pennsylvania with guaranteed human oversight. The push comes as nearly 70 percent of US government agencies already use generative AI, per an Ernst and Young survey.

Former OpenAI researcher Will DePue says he found technical evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group trained models on Nvidia Blackwell GPUs despite US export controls, estimating the hardware behind five runs cost over half a billion dollars.

Anthropic's IPO is six times oversubscribed and could raise up to $64.4 billion, with a target valuation near $1.4 trillion. Nvidia is buying $10 billion in stock through a parallel private placement.

OpenAI has launched two cybersecurity products under its Daybreak brand: Daybreak Red Team, in Early Access, which uses AI agents to probe web apps for vulnerabilities, and Daybreak Defense Factory, which scans codebases and generates patch pull requests. Both run on OpenAI's GPT-6 Cyber model.

Google released Gemini 4 Hydrogen, its replacement for Gemini 3.8 Flash, which scores 51 on the Artificial Analysis Intelligence Index at the default thinking level — edging out the larger Gemini 4 Argon at 50 and reaching 54 at a new xhigh tier. It costs $1.50 per million input tokens and $7.50 per million output tokens.

OpenAI's safety team says the declining legibility of frontier models' chains of thought stems from four training steps, including architecture choices and reinforcement learning, framing it as a tradeoff around capability and cost rather than an inevitability.