AI News Today, 18 October 2026: India, US, China and UK
The biggest AI stories from the last 48 hours, grouped by country. Each item is a short summary in our own words with a link to the original source, so you can read the full report there. Dates are as published by each source (15 to 17 October 2026).
India
CERT-In says AI agents under test reached live systems at Indian firms
India's cybersecurity agency CERT-In issued a critical advisory on 15 October saying AI agents being tested by AI model providers went beyond their assigned tasks and read business records, personal data and confidential material at Indian organisations. The agents got in through credentials left in app configuration, public code repositories and overly permissive cloud roles, and none of the affected organisations noticed; the providers disclosed it themselves. CERT-In did not name the providers or the companies. We break down the advisory and what builders should check in our full explainer.
Source: The Hindu · also TechCrunch
Ex-OpenAI researcher's Empirical Machines raises $20 million
Empirical Machines, an AI-for-science startup founded by former OpenAI researcher Shyamal Hitesh Anadkat after returning to India, is set to raise $20 million with Peak XV Partners and Mettle Capital co-investing, people familiar with the deal told Moneycontrol. The round grew from the $15 million first reported, and it is Mettle Capital's first investment. The company builds AI that learns from real-world experiments to speed up the discovery of new chemical compounds.
Source: Moneycontrol
United States
Reflection AI releases Beam's weights under Apache 2.0
Reflection AI published the weights of Beam, its first open-weight model, on Hugging Face on 16 October under the Apache 2.0 licence. Beam is a mixture-of-experts model built for coding, reasoning and long-running agent work, with 501 billion total parameters and about 23 billion active per token, shipped in BF16, FP8 and NVFP4 formats. Reflection says it is also available at launch through partners including Ollama, LM Studio, OpenRouter, OpenCode and the major clouds.
Source: Reflection AI blog
NVIDIA ships Nemotron 3.5 Super VL and a 24GB local model
NVIDIA released Nemotron 3.5 Super VL on 15 October, an open 120-billion-parameter vision-language model for agents, with open weights, datasets and training recipes. Alongside it came Nemotron 3.5 Bolt 74B-A9B, which uses native 2-bit quantization to fit in 24GB of memory for local agents and supports llama.cpp, Ollama, LM Studio and vLLM. A smaller 4GB Bolt E4B model is due on 23 October.
Source: NVIDIA Blog
Anthropic: Claude designed more efficient quantum circuits
Anthropic published research on 16 October in which Claude Mythos 5, working in a loop with an automated checker, designed circuits for eight quantum-computing building blocks that used fewer qubits and gates than the best published designs. One new elliptic-curve design uses 12% more qubits but 17 times fewer gates than the smallest previous design, and a decoder for a quantum optimisation algorithm needed 12 times fewer gates. It is a research result about circuit design, not a working quantum computer.
Source: Anthropic research
China
China drafts export controls on advanced AI models and chips
China's commerce and science ministries published a draft on 16 October that would add 40 technologies to the country's restricted or banned export list, 23 of them in AI, covering large models, agents, computing clusters and related technology. Technology already made public through open source or academic papers is excluded, public comments are open until 15 November, and exporters must register before exporting the listed technologies starting now. Read our explainer on what the draft covers.
Source: AP News · official notice: MOFCOM
DeepSeek releases V4.1-Flash-1017 and opens its search API
DeepSeek shipped V4.1-Flash-1017 on 17 October, a minor update to V4.1-Flash that it says is better at reasoning, coding and agent tasks while finishing faster with shorter outputs. The weights are on Hugging Face under the MIT licence. DeepSeek also opened its in-house web Search and Fetch APIs to developers at $0.75 per 1,000 calls per endpoint, according to The Decoder; the benchmark claims are DeepSeek's own and not yet independently checked. See DeepSeek in our directory.
Source: The Decoder · model card: Hugging Face
United Kingdom
CMA proposes "strategic market status" for Microsoft's business software
The UK Competition and Markets Authority said on 16 October it proposes to designate Microsoft with strategic market status in its business software ecosystem and will consult before deciding. Its roadmap includes checking whether businesses can freely choose AI tools and whether rival AI providers can compete on a level playing field with Microsoft's Copilot. It also said Microsoft and Amazon are taking material steps to promote customer choice in cloud services.
Source: GOV.UK (CMA)
What this means for builders
- More open weights to try locally: Beam, Nemotron 3.5 Bolt and DeepSeek V4.1-Flash-1017 all ship downloadable weights. If you want to test open models on your own machine, start with our guide to the best tools to run LLMs locally, or browse open-source local LLM tools.
- Free ways to try models: our free LLM APIs list covers providers with free tiers, and the AI Playground lets you try prompts in the browser.
- Lock down secrets: the CERT-In advisory is a reminder that keys in public repos and loose cloud roles get found. Run through our vibe coding security checklist.