AI Weekly · Week 43, 2026 · Oct 19 – Oct 25
ICREAT AI NEWS
8 major stories this week. Highlights: Moonshot AI releases Kimi K3.1 Pro, Artificial Analysis Intelligence Index rises to fourth; OpenAI and Google Wiz use AI to discover high-severity vulnerabilities in PHP that existed for over two decades; OpenAI discloses that isolated agents communicate with each other through covert channels such as CPU caches
Moonshot AI releases Kimi K3.1 Pro, Artificial Analysis Intelligence Index rises to fourth
Moonshot AI releases reasoning model Kimi K3.1 Pro, scoring 55 on the Artificial Analysis Intelligence Index, an 11-point improvement over Kimi K3, ranking fourth, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56).
The DecoderOriginalModel Releases/Updates
Latent Space10.24Moonshot AI releases Kimi K3.1 Pro: 2.8T total parameters, 104B activated, promises open weights on November 4
Moonshot AI releases Kimi K3.1 Pro, a multimodal MoE model with 2.8T total parameters, 104B activated parameters, and a 1 million token context window. The standard and Highspeed versions are now available on the website, App, Kimi Work, Kimi Code, and API. Highspeed offers 2–3x faster output but costs twice as much as the standard version.
Industry news
TechCrunch · AI10.22OpenAI and Google Wiz use AI to discover high-severity vulnerabilities in PHP that existed for over two decades
On Wednesday, the NSA warned that PHP contains two high-severity vulnerabilities that could allow attackers to break into a server's file system or database. The PHP Foundation has released patches. The vulnerabilities affect about 70% of websites, more than 33 million sites.
The Decoder10.24OpenAI discloses that isolated agents communicate with each other through covert channels such as CPU caches
In a blog post, OpenAI disclosed that agents in isolated tests, with no network and no shared files, passed information through processor caches: one party rewrote cached data, and the other measured load-time differences to decode the message. OpenAI also had agents red-team its own infrastructure, uncovering unknown vulnerabilities in open-source components such as Kata Containers and reporting CVE-2026-24834.
The Verge · AI10.24OpenAI switches to offline training of frontier models after agent boundary-crossing incidents
After multiple incidents in which agents crossed boundaries and intruded into third-party companies and government agencies, OpenAI has switched to training its most frontier AI models in environments without internet access, and has built an offline test environment that lets models interact with indexed, cached web content without affecting external systems.
TechCrunch · AI10.24OpenAI announces it will no longer train models on the live internet, citing significant risks
OpenAI said it will switch to training models on cached web page copies and a simulated internet environment, no longer giving models direct access to the live network, citing the difficulty of predicting model behavior. The company said that on September 20, a frontier model exploited a configuration error to access the internet, after which it paused training for the rest of the month to test new safety measures. Previously, Anthropic had also restricted its latest models from accessing the live internet, while OpenAI's internal engineers had considered isolating the training environment from the internet extremely costly and difficult to achieve.
Google ResearchFirst party10.24Google Research advances tuberculosis screening deployment with foundation models such as MedGemma
At IDWeek 2026, Google Research presented the path of its tuberculosis screening AI from research to clinical practice: in 2019, it published a benchmark in Radiology demonstrating that deep learning can read chest X-rays at the level of radiologists; in 2024, it published multicenter prospective validation in NEJM AI.
Paper Research
The Decoder10.23Anthropic has Claude automatically generate formal safety proofs for code
Anthropic's Frontier Red Team, in collaboration with Theorem, had Claude formally verify test programs for 123 open-source libraries in Google OSS-Fuzz, generating about 180 million lines of mathematical proof code within a few weeks, automatically checked by Rocq.