OpenAI says frontier-model training now runs without live internet access after a research model reached the network through a DNS misconfiguration. Training resumed October 1 after infrastructure changes. Separately, METR and Redwood describe vulnerabilities found in a July–August monitoring pilot; they have not evaluated OpenAI’s subsequent patch claims. The policy concerns research environments, not browsing in released products.
Previously 2026-10-22 · METR and Redwood describe flaws in the historical monitoring system; their pilot does not verify current coverage or subsequent patches.History −
OpenAI made its Decisions API generally available for bounded decisions such as classifying inputs and choosing a workflow’s next action. Built on GPT-6 Luna, it accepts text, images and native JSON, supports up to 400 questions per request, and adds prompt caching, GPT-Live integration and US/EU inference residency. Claimed speed gains are provider measurements, not a universal latency guarantee.
Anthropic is donating the Agent Skills standard, name, domain and GitHub organization to the Linux Foundation’s Agentic AI Foundation. Skills package task instructions and supporting material for agents. Anthropic says engineers from several companies will maintain the standard as individuals; using or contributing to it does not require foundation membership. Versioned releases and security guidance are proposed next steps.
Anthropic opened a research preview of Claude Duet for US Pro and Max subscribers, initially through a waitlist. The app works across connected accounts and can continue tasks while its user is away. Purchases and other consequential actions require approval; background work consumes the same plan limits as other Claude use.
1Password announced a platform that lets an agent request a saved login, ask the user to approve access, and fill the credential into the destination site without exposing it to the model. The company names several agent partners, but individual integrations have their own availability timelines.
Block has introduced Dessa, a speech-to-speech model it says was trained from human conversations without a text model. The company reports local operation on an M4 Max MacBook Pro or newer and is inviting a limited group of safety researchers to test it. This is a research preview, not confirmed general availability.
Legora announced a $550 million Series E led by Atreides Management at an $8.8 billion post-money valuation. It plans to expand in North America and other markets and develop its platform for multi-step legal work. This is a separate round from the $550 million Series D announced in March; equal amounts do not make the transactions duplicates.
Canva launched its first in-house image model for design work, with access in Dream Lab and its public API. Craft separates composition planning from rendering. Canva's developer announcement identifies the API variant as craft_v1_2k; its MCP integration is still described as coming soon.
METTA combines within-episode neural memory with compact summaries across episodes, according to its preprint abstract. The authors report adaptation results on held-out reinforcement-learning tasks. This is research on meta-RL agents, rather than evidence that a general-purpose language-model assistant has acquired reliable long-term memory.
The authors propose choosing which partially generated solutions to continue by directly optimizing pass-at-k. Their abstract reports tests across open-weight reasoning models. The method tries to retain promising but varied reasoning paths, rather than optimizing quality and diversity separately; these are preprint claims, not independently reproduced results.
The authors compare trained classifiers and prompted language models under explicit costs for correct routing, human triage and mistakes. Their abstract reports that abstention policies can reduce routing cost, while better calibration alone need not improve a model's ability to distinguish correct answers. The findings are limited to the tested dataset, prompts and illustrative costs.
