A quiet weekend for new models; the main story is Anthropic's IPO, whose order book is reported to be about six times covered ahead of pricing expected next Wednesday. Cursor confirmed a breach that led to malicious copies of the Claude Code extension being distributed, a reminder of supply-chain risk in agentic coding tools. Google published a batch of research results from Antigravity's multi-agent /teamwork, and OpenAI disbanded its stand-alone robotics team, saying general-purpose models can now run robots.
According to Reuters (citing two sources, as carried by Calcalist and NewsBytes on Nov 14), by the end of Thursday orders for Anthropic's roughly $64 billion IPO were about six times the shares on offer, up from about twice after the first roadshow day. Pricing is expected on Wednesday (Nov 18) when the roadshow ends; investors still question how long Anthropic can keep its lead over OpenAI and Google.
Fortune reported on Nov 13 that OpenAI confirmed it disbanded its stand-alone robotics team: researchers moved into the main research org under Mark Chen, while applied and hardware staff moved to Greg Brockman's Applied org. The rationale is that general-purpose models such as GPT-6 Astra already run robots directly (74.8% success on an Nvidia simulated manipulation benchmark versus 51% for the best specialized model); OpenAI says this is not a retreat from robotics.
On Nov 14, Cursor confirmed attackers took an access key from infrastructure provider Fireworks AI, copied coding-agent traces (some with unredacted API keys), and on Oct 29 placed two fake Claude Code extension versions, 2.1.312 and 2.1.313, in its marketplace cache for about 11 hours, attempting to install SSH backdoors. Cursor estimates up to 5,000 users may have received the malicious builds; Anthropic's official extension and Open VSX were not affected.
A Google Research post on Nov 13 says /teamwork, Antigravity's multi-agent mode, helped prove the 2013 Courtade–Kumar conjecture with a Lean formalization, rewrote the libgav1 AV1 decoder in Rust running 20–30% faster than the C++ original, and sped up TPU kernels 4.09x. Gemini 4 with /teamwork and /boost is available to external users, and Google is offering 10,000 academic researchers six months on its top usage tier.
On Nov 13, ggml-org released libllama and libggml 1.0.0 and moved to semantic versioning, so major version bumps will signal breaking C API changes. After 3.5 years the project supports more than 150 architecture variants, ships packages for Linux, macOS, Windows, Android and iOS, plus desktop apps built with Hugging Face, cementing its role as a base layer for running LLMs locally.
VnExpress reported on Nov 13 that Nguyen Tran Thanh Lam, a third-year NTU (Singapore) student, and collaborators used Claude Opus 5 to help find CVE-2026-125742, a heap out-of-bounds write when 7-Zip extracts multi-part CAB files that enables remote code execution (CVSS v4 8.5). It is fixed in 7-Zip 26.06 and later; a concrete case of LLMs aiding real vulnerability research.
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗