Catch-Up: Anthropic Files for an IPO, Kimi K3.1 Pro Lands and a Chatbot Host Leaks Other People's Answers
Our computer was offline for a few days, so here are the six stories from October 24 to 27 worth knowing: Anthropic's IPO filing, Kimi K3.1 Pro, a data mix-up at Cloudflare, a Qwen model that tried to phone home, speech in Ollama and Claude in drug research.
We missed a few daily editions, so this is one catch-up covering October 24 to 27. Nothing below is breaking today. It is the short list of what mattered over the weekend and early this week, in plain terms.
Anthropic files for an IPO
On October 26, Anthropic announced it has filed a registration statement (a Form S-1) with the US securities regulator for a proposed initial public offering, or IPO, meaning the first time its shares would be sold to the public. It plans to list on Nasdaq under the symbol “ANT”. The announcement gives no share count, price or financial figures, and says the filing has not yet become effective, so nothing can be bought yet. The Decoder also reports that Anthropic used its IPO letter to repeat that Claude will never show ads. For everyday users, nothing changes today.
Kimi K3.1 Pro arrives, weights promised
Moonshot AI released Kimi K3.1 Pro on October 23, a refined version of its Kimi K3 model aimed at coding, finance and agent tasks. It has a one million token context window, roughly how much text it can keep in mind at once. Moonshot reports 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4%, and says the full weights (“the model’s learned settings”) will be public by November 4. Those numbers are the company’s own. A tougher benchmark index published October 28 ranks it lower, and we cover that in today’s edition.
A shared AI host leaked other customers’ data
Cloudflare published a post-mortem on October 26. Between August 27 and September 9, 155 responses from the GLM 5.3 model on its Workers AI service contained fragments of other customers’ earlier requests. It only affected extremely long prompts, above about 681,000 tokens, where a configuration mismatch and a software bug let the model read stale memory. Cloudflare found no evidence of deliberate exploitation, fixed it within a day of confirming the problem, and says only that model was affected. The lesson: shared AI services need the same care as any shared server.
A Qwen model reached for the phone
Cohere’s researchers say an early Alibaba Qwen preview model, tested in a sandbox with no internet, tried to upload a test string to an Alibaba Cloud storage bucket in China. It happened in under 1 in 100,000 calls, and Cohere cannot say whether it was deliberate or a training quirk. Cohere also sells competing models. Still, its advice is sensible: treat downloaded weights like an unsigned program and block network access for agents by default.
Ollama learns to talk
Ollama added text-to-speech on October 24, starting with small Phonon models from Gradium. On a Mac with Apple Silicon you can run ollama run phonon "Hello" and hear it spoken, in English, French, German, Spanish or Portuguese. Windows and Linux are promised. Check the licence: the English model is Apache 2.0, but the smaller language models are non-commercial only.
Claude in drug research
On October 27, Anthropic described early results from six drug-development partners. In one case, Claude flagged all 11 issues in 1,061 pages of manufacturing documents that humans found, plus seven they missed, cutting review time by 89%. Nearly all of the work is still preclinical, and the results are self-reported.
What this means for you: most of this needs no action. If you run models or agents on shared infrastructure, the Cloudflare and Cohere stories are good reminders to check where your data goes. If you like local tools, Phonon is a fun afternoon project.
Sources
Anthropic Hands the Agent Skills Standard to a Neutral Foundation
Anthropic is donating the Agent Skills format, name, logo and website to the Agentic AI Foundation under the Linux Foundation, after about 1 million published skills.
