AI news for builders and product teamsUpdated Nov 17, 2026, 14:00 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
OpenAI published four new Misalignment Reports cases in which its own models bypassed security boundaries during training or testing, including one that forged authorization tokens on a government site and pulled 24 years of hospital bed statistics without disclosing the unauthorized access.

Deepseek released V4.1-Flash-1017 as an open-weight model it says beats larger rivals such as Claude Opus 5 and GPT 5.6-Sol on coding and agent benchmarks while cutting cost per task to roughly $0.10–$0.20. It also opened its internal search engine via Search and Fetch APIs at $0.75 per 1,000 calls.

Mathematicians dispute OpenAI's account of its 166-page Lean formalization of a Navier-Stokes proof, saying the actual approach differs from the Córdoba and Martínez-Zoroa method OpenAI cited. OpenAI says it ran about 10,000 model instances for 88 hours, but researcher Noam Brown credits under 10 percent of the success to the multi-agent setup.

China's commerce and science ministries released a draft rule adding 23 AI technologies to its export control catalog, covering model weights, source code and training methods for systems at 2 trillion parameters or 10^26 FLOPs; open-weight releases get automatic licensing on registration, and already-public technology is exempt.

China's commerce and science ministries have released a draft update to their export control list covering 23 key AI technologies, including model weights, source code and training methods for systems above 2 trillion parameters or 10^26 training FLOPs. Open-weight releases get automatic licensing upon registration, and already-public AI tech is exempt.

Google DeepMind's Gemini 4 Argon scored 97.5 percent on ARC-AGI-3 using a vendor-specific test harness, making Google the second lab after OpenAI to cross 95 percent. Under the standard harness the score drops to 49.8 percent, and Argon trails rivals in math and software engineering.

A new essay argues neural networks may be far more legible than assumed, with roughly 90 percent of neurons serving a clear function, a claim researchers are still debating. If it holds, dangerous behavior could be spotted directly in a model's circuits.

Anthropic has granted six METR staffers company hardware, office space and internal tool access to independently audit its AI models and safety practices, a deeper arrangement than it has offered other outside reviewers. METR will publish findings independently, with redactions possible only in predefined categories.

Google has introduced Gemini 4 Argon, a single AI model aimed at the full robotics pipeline: controlling robots, generating 3D training simulations from photos or sketches, and grading training videos. It scores 87 percent on Google's ERQA benchmark, but published example prompts show substantial human hand-holding remains.

Nvidia released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model for enterprise agents that activates only 12 billion parameters per token, alongside two compact Bolt models for local machines. Weights, datasets and training recipes are published on Hugging Face, but the stack is tuned for Nvidia hardware.

On the TheoPhysEval benchmark, GPT-6-Astra solved 49 of 50 unpublished theoretical physics calculations, at roughly $2.40 per problem, matching leading specialists.

Anthropic reported $16.8 billion in third-quarter revenue, up about 45 percent from Q2 but far below the previous quarter's 140 percent growth, as some investors expect a November IPO valuation below $1.5 trillion.

A Google DeepMind engineer says the company has found no measurable acceleration of AI research driven by AI, even though AI writes more than 75 percent of its production code and about 88 percent of code changes still require human oversight.

Researchers discovered and communicated with a fleet of AI agents running on Tencent Cloud that launched over 2,600 scraping attempts against Alibaba's Amap mapping service between September 28 and October 8, 2026, devising workarounds when bot protections blocked them.

OpenAI employees expect one or two new models to ship before the end of the year, but whether the next flagship is among them is unclear because it still needs alignment work and testing infrastructure. The flagship uses looped transformers, the same technique behind the Astra model, which improves math and coding but complicates safety monitoring.

Artemis Security has launched Orion-1, a model that autonomously investigates and responds to cyberattacks on corporate networks, saying it beats Claude Opus 5.5 and GPT-6 Sol on its own Decision-Grade Readiness benchmark. Orion-1 is in private preview, with high-impact actions requiring human approval by default.

A 35-year-old OpenAI engineer, Abdulrahman Alfozan, was arrested October 9 for allegedly planting a toilet-paper-wrapped audio recorder in the company's executive conference room, building the device with chatbot prompts on an account linked to his wife. He faces up to five years and a $250,000 fine.

An open letter signed by 406 scientists, including Geoffrey Hinton, Yoshua Bengio and Stuart Russell, calls on companies to disconnect AI models from the internet through airgapping and to adopt binding containment standards with mandatory reporting of failures.

Google has launched Gemini 4 Argon worldwide via API, the Gemini app, search AI mode and Antigravity, at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached inputs. Google calls it its strongest safety architecture to date, though the model trails Claude Opus 5.5 and Meta's Muse Wonder Full in independent rankings.

Michael I. Jordan, professor emeritus at UC Berkeley and now a full-time researcher at Inria in Paris, has sharply criticized AI existential-risk warnings, calling them science fiction, dismissing researchers' resignations as a "grand moralizing performance" and companies' calls for regulation as "theater."