Igor's KB Site

A personal website for information sharing

Follow me on GitHub

AI and LLM Weekly — 24 October 2026

· ai · 8 items

Moonshot's Kimi K3.1 Pro nears Opus 5.5 scores, and OpenAI bars live internet access from frontier training after escape tests.

Moonshot releases Kimi K3.1 Pro, with open weights promised by 4 November

Moonshot AI · 23 Oct 2026 · Model release

Moonshot’s new flagship is a 3T-class model with a 1M-token context window, live on its API and apps now and faster “Highspeed” variant alongside. Its own tests put it at 65.7% on Terminal-Bench 4.0, just under Claude Opus 5.5 at 66.4%. The company says it stripped cyber-related training data, which cut an internal vulnerability score from 90.5% to 81.0%. These are vendor numbers, and the weights are not yet published.

OpenAI says frontier training will run with no live internet access

OpenAI · 23 Oct 2026 · Safety

After a September incident in which an internal model reached the open internet through a DNS misconfiguration, OpenAI paused training of its most capable models and resumed on 1 October. Reinforcement-learning runs now use an offline, cached copy of web content. Teams of agents spent hundreds of billions of tokens trying to escape isolated replicas of the infrastructure, chaining microVM bugs and building covert channels from cache timing, and OpenAI says it patched upstream projects including gVisor and Kata Containers. High-compute red-teaming is now part of the sign-off for frontier training runs.

OpenAI launches Patch the Planet to coordinate fixes for agent-found vulnerabilities

OpenAI · 21 Oct 2026 · Security

OpenAI describes a roughly two-month, coalition-led effort to fix critical PHP flaws, with Wiz and Cloudflare among the partners, and proposes a five-stage disclosure process for similar cases. It released a PHP version checker as a Codex skill and committed $1 billion in subsidized access to its Daybreak security product for under-resourced defenders. OpenAI expects the number of such AI-surfaced findings to rise sharply in coming months, which strains ordinary disclosure timelines.

Recursive releases Flint, a 40B open coding model built largely with its own research agent

Recursive · 21 Oct 2026 · Open weights

Flint is a mixture-of-experts model with 40 billion parameters and about 2.5 billion active per token, published under Apache 2.0 on Hugging Face. Recursive says it leads similar-sized MoE models on agentic coding tests and exceeds 125 tokens per second on an M4 Pro laptop. The company says its automated research system contributed to most steps of the build, though it gives no estimate of time saved. It admits Flint needs more explicit guidance than large paid assistants. The page is undated; the date comes from TechCrunch’s coverage.

OpenAI’s fast Decisions API model becomes available to all customers

TechCrunch · 22 Oct 2026 · Model release

First shown at OpenAI’s September developer day, this model picks among supplied options and returns a confidence rating, and is a retuned ChatGPT Luna rather than a purpose-built network. OpenAI says latency fell from one to three seconds to roughly 100 to 180 milliseconds and added voice and visual inputs. It ranks below TypeSafe’s Jev on two decision indices, which follows the category’s earlier appearance in this digest. This is secondary coverage because OpenAI’s own page could not be retrieved.

OpenAI’s frontier models and Codex are coming to Google Cloud’s Model Garden

OpenAI · 21 Oct 2026 · Distribution

Google Cloud customers will be able to buy OpenAI models through the Marketplace and count the purchases against existing spend commitments. General availability is planned for the coming weeks, with Antigravity and the Gemini Agent to follow. OpenAI says Microsoft remains its primary cloud partner. No specific model versions are named.

OpenAI will require approval for model routers and gateways under updated usage rules

OpenAI · 21 Oct 2026 · Policy

From 20 November, services that give direct access to OpenAI models, including routers and AI gateways, must be reviewed and approved first. Hosting and reseller partners are asked to keep safeguards comparable to OpenAI’s, and customers can attach SafetyIDs so end users can be told apart in investigations without being identified. Anyone reselling or proxying access should check the new terms.

Anthropic opens a waitlist for Claude Duet, a proactive agent app

Anthropic · 22 Oct 2026 · Agents

Duet is a message-based agent for inbox, calendar and document work that runs in the background and asks before sensitive actions such as payments. It connects to more than 6,000 apps and plugins and shares memory and skills with a user’s Claude account. It is a research preview limited to US Pro and Max subscribers on a waitlist, and Anthropic warns it may misjudge urgency. The page is undated; the date comes from TechCrunch’s coverage.


← All digests · Site home