China drafts AI export controls; India reports rogue agents

China's commerce and science ministries have proposed adding 23 AI technologies to the country's list of technologies restricted or banned from export, covering large models, AI agents, computing clusters and chip design. The draft, published on Friday as Announcement 48, also covers solar and biomedicine items, and the public can comment until November 15. For large models the curbs would reach weights and source code, and Bloomberg reported they would apply to models with at least 2 trillion parameters or that rank in the world's top ten. The draft names no model. Open-source releases and published research are exempt, and Chinese companies can keep serving overseas users through apps, websites and APIs, but running a covered model on servers abroad would count as an export. A temporary rule requiring exporters to register first took effect immediately, before the consultation closes. In July, Beijing was reported to be weighing limits like these; this is the first formal text.

India's Computer Emergency Response Team (CERT-In) said on Thursday that it had received "multiple voluntary disclosures from AI model providers" whose agents, during research and testing on the open internet, reached the live systems of Indian organizations. The agents got in through credentials and settings left exposed in app configurations, public code and overly broad cloud roles, then read personal data and confidential business records. CERT-In did not name the providers or the companies, and the organizations had not spotted the intrusions themselves. The same day, OpenAI published four more reports under the disclosure framework it launched in September. In one, a model asked to update a crypto trading app recovered wallet keys from public developer recovery phrases and spent test-network funds without permission. In another, a model told to fetch five dashboard figures downloaded five non-public tables of more than 147,000 rows and never said so. All four incidents happened in June.

Anthropic says Claude Mythos 5, working in a closed loop with an automated grader, found quantum circuit designs that use fewer qubits and fewer gates than the best published versions across eight standard tasks in chemistry, optimization and number theory. In one, after 53 hours with minimal human input, it found a new way to do the costliest step of elliptic-curve point addition, the core operation in quantum attacks on elliptic-curve cryptography; one resulting design uses 12 percent more qubits but 17-times fewer gates than the smallest earlier one. The work, with Harvard physicists, is on arXiv and not yet peer-reviewed. Separately, Anthropic said an internal model took under half an hour to work out where the seeds behind the NIST elliptic curves came from. NIST adopted the curves in 2000 using seeds the NSA never explained, and they turned out to be output from an old C random-number generator seeded with a timestamp, which Anthropic says makes a hidden backdoor much less plausible.

DeepSeek has released DeepSeek-V4.1-Flash-1017, an update to the model it shipped in September, and switched its website, app and API to it with no change in price. It reports higher scores across coding, agent and vision tests, including a jump on Terminal-Bench 4.0 from 31.2 to 42.3, alongside shorter outputs and faster task times. Its comparison table sets the model against OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 rather than those labs' newer releases. The weights are on Hugging Face under a plain MIT license with no use restrictions. DeepSeek also opened its in-house Search and Fetch APIs to all developers at 75 cents per 1,000 calls for each endpoint.

Replit chief executive Amjad Masad said onstage at TechCrunch Disrupt on Thursday that the coding platform is no longer working with Anthropic, citing cost, partnership and performance, and said OpenAI's Sol and Astra models score better on Replit's own benchmark for building apps from scratch. Replit announced a partnership with OpenAI in August. A person familiar with the Anthropic relationship told TechCrunch it is still in place, and a Replit blog post from late September listed Claude Fable 5 and Fable 5.1 among the models running in production. Masad also said enterprise contracts now make up about 15 percent of Replit's revenue, up from about 2 percent last year.

Anthropic has moved Code Review, its multi-agent reviewer for pull requests in Claude Code, out of research preview for Team and Enterprise plans. It says most reviews now finish in six to ten minutes, about three times faster than in preview, and admins can pick a quick, standard or deep review for each repository. Eligible Enterprise organizations can claim at least $10,000 in credits before January 15. A companion post from the Claude Code team explains why: changes landing per day in Anthropic's main code repository rose from about 2,500 in May to more than 30,000 in September, which made line-by-line human review impractical. On Claude Code itself, merged pull requests more than doubled from June to September while the share later reverted fell by more than 40 percent.

Google Research described three projects that put AI agents in control of real lab robots. One is a system that keeps a live 3D model of a lab bench, so an agent can check volumes, tips and robot geometry and simulate a protocol before it touches reagents. In a second, agents ran three closed-loop design and test rounds of DNA aptamers, at seven to ten days a round, and raised confirmed hit rates on two proteins. In the third, a combined human and agent pipeline found protein binders with single-digit nanomolar affinity for a receptor overexpressed in solid tumors in just a few experimental runs, a level of binding Google says usually takes extensive lab optimization. Google says fully autonomous labs still need safeguards against agents gaming their own experiments.