2026-10-01 digest
2 items on this page: 1 with their own page, 1 short.
A digest puts the short items from one day or one week on a single page. Opened alone, those briefs are often a headline and a sentence, which is a poor fit for a standalone article. Items that meet the importance threshold keep their own pages and are listed here by title.
Items with their own page
Google introduces Gemini 4 Argon, starting with trusted cyber defenders
Short items
OpenAI says it disrupted a coordinated campaign to extract protected model reasoning
【What it is】
On September 30, 2026, OpenAI said it identified and disrupted a coordinated campaign to extract protected reasoning from its models. Earliest activity was in the first week of July. The post says it began July 1 and was fully disrupted by July 28.
【Verifiable】
It calls this adversarial distillation: unauthorized use of one model’s outputs or reasoning to train or copy another. Protected reasoning is the model’s internal working record. Operators, OpenAI says, did not break encryption, breach a database, or read stored chats. They manipulated chats so reasoning became visible, at scale, against the terms—for example by pasting encrypted reasoning into another chat and asking the model to decrypt it. OpenAI says it confirmed cross-model and conversation-compaction paths reported by researchers linked at https://arxiv.org/abs/2608.09867.
July 24–25: 16,000 requests from over 4,000 users. The footnote says these figures are attempts, not necessarily successful extractions. Related prompt patterns spanned a cluster of more than 15,000 users. OpenAI says it is unclear every operator was one actor, and ties a core cluster to individuals associated with Moonshot AI, maker of Kimi.
It reports account bans or limits, stronger signup and monitoring, closing a replay path for encrypted reasoning someone already held, holding streams that might leak reasoning, work with third parties, and sharing via the Frontier Model Forum and government channels. More mitigation is still underway.
【Our read】
This is OpenAI’s account of a terms-of-service campaign it says it stopped, not a broken cipher or direct access to stored chats. Scale on the page is requests and a user cluster; the footnote marks those figures as attempts. Attribution is one core cluster of people tied to Moonshot AI, not every operator. The safety point is OpenAI’s own reason in the post.
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy