Igor's KB Site

A personal website for information sharing

Follow me on GitHub

AI and LLM Weekly — 7 November 2026

· ai · 7 items

OpenAI rates GPT-6 Cyber as Critical for cyber risk and restricts access; open-weight releases from Moonshot and Tencent headline seven items.

OpenAI launches GPT-6 Cyber and rates it Critical for cyber risk

DataCamp · 5 Nov 2026 · Model release

OpenAI moved its cybersecurity rating from High to Critical for this model, up from GPT-5.6 Cyber, and is limiting use to approved customers who pass identity checks and attest that their testing is authorized. Reported scores include 48.5% on ExploitGym against 32.1% for its predecessor, and 81 new high-severity Firefox bugs, many already fixed by Mozilla. All figures are OpenAI’s own. This links to secondary coverage because OpenAI’s announcement page could not be retrieved.

OpenAI says one autonomous run produced 948 working Linux root exploits

AI Weekly · 4 Nov 2026 · Security

OpenAI describes a pipeline in which an unnamed upcoming cyber model scanned the latest Linux kernel, validated 1,724 candidate flaws on hardened builds and delivered 948 local privilege-escalation exploits, each with a candidate patch. The run cost about $1.21 million in API spend, roughly $1,279 per flaw, against 166 flaws from a July attempt with an older model. Ten days after the scan, 61 issues were patched upstream and 299 were being reported publicly. These are OpenAI’s numbers, relayed second-hand because OpenAI’s own page could not be retrieved.

Moonshot releases Kimi K3.1 Pro weights, a 2.8-trillion-parameter multimodal model

Moonshot AI · 5 Nov 2026 · Open weights

This delivers the open weights promised when the model launched in October. It is a mixture-of-experts system with about 104 billion parameters active per token and a context of roughly one million tokens, taking text and image input. The license is a custom Kimi K3.1 Pro License whose terms the page does not summarize. Moonshot’s benchmark results are self-reported on partly internal tests, and the model trails GPT-6 Astra on some reasoning evaluations.

Thinking Machines launches Inkling 2 Small and promises Apache weights next week

Thinking Machines Lab · 5 Nov 2026 · Model release

The 594-billion-parameter mixture-of-experts model, with 17 billion active, is available now through the company’s Tinker platform with closed weights. The lab says it intends to publish full weights under Apache 2.0 next week, after external safety testing and red-teaming. It reports a score of 31 on the Artificial Analysis Intelligence Index, well above the lab’s earlier models, though it still trails closed models like GPT-6.1 Sol on the lab’s own alignment evaluations.

Tencent releases Hy4 with vision and computer use, following August’s preview

Tencent · 4 Nov 2026 · Model release

The full release adds image understanding and the ability to operate command-line and graphical interfaces, which the August preview lacked. Tencent reports its internal hallucination rate fell from 12.95% to 7.2%, and pricing stays near $0.90 per million input tokens and $2.70 per million output. Weights and system cards are promised only “in the coming days”, so the open-source label is still a pledge.

Anthropic open-sources FlashProve, a way to prove which model produced a reply

Anthropic · 2 Nov 2026 · Research

The prototype lets a provider commit to a model’s weights and attach a zero-knowledge proof to each response, so anyone can check which model generated it without seeing the weights. On a 400-billion-parameter model it proves about 110 tokens per second on one H100, but costs roughly 70 to 95 times the GPU time of ordinary serving and covers only 1,024-token contexts under greedy decoding. Code and Lean proofs are Apache 2.0, and Anthropic calls it a research prototype.

Claude helps determine the first 3D structure of a bacterial vault particle

Anthropic · 4 Nov 2026 · Science

Anthropic says Claude searched a database of predicted proteins, designed DNA constructs, suggested purification methods and ran cryo-EM processing in a lab project that solved the structure of a phage-derived vault with 96 protein copies. The vault’s biological function remains unknown, and the particles have only been made in E. coli. Anthropic also reports Claude’s mistakes, including an additive that cost about two weeks. The results are in a bioRxiv preprint and have not been peer reviewed.


← All digests · Site home