The Day in AI

• Anthropic launched Claude Duet on 21 October, a research preview agent for professionals and small businesses that messages users only when something needs attention.
• MemTensor open-sourced MemT 1.0 Flash and Nano, models with a fixed-size trainable memory that it reports runs about 4x faster at 256K tokens of history.
• OpenAI reportedly had a Calif researcher using GPT-5.6 Sol found PHP flaws named PHP2Shell, and Wiz found them independently, according to Latent Space.
• Canva released Canva Craft, its first in-house image model, trained on its own design templates and available to all users in Dream Lab.
• Instinct, an AI agent startup, reportedly raised $1 billion at a $10 billion valuation, according to Crypto Briefing.
• METR reported that a pilot red-team of OpenAI's internal agent monitoring found an exploit that could have let agents hide hundreds of recent actions.

Claude Duet is a research preview aimed at professionals and small business owners. Each morning it sends a brief on key items across the user's connected apps, and during the day it messages only when something needs attention. It can schedule meetings, build dashboards, make forecasts and build web apps. It uses the memory, skills and connected apps a user already has in Claude, and it runs in its own virtual machine with a browser. Anthropic says it asks for approval before big steps such as payments or deletions. A new dedicated classifier and safety system prompt reject requests to share the most sensitive information, such as credit card numbers.

MemTensor open-sourced two models with memory built into the network. Most assistants keep the whole conversation in the context window, so every turn gets more expensive. MemT instead writes each new interaction into a fixed-size internal state and reads from it later. MemTensor reports that MemT-1.0-Flash has 288 billion total parameters with 17 billion active, and Nano has 36 billion with 3 billion active. At 256K tokens of history, the company reports end-to-end inference 3.99x faster than full context, and a memory state of about 40 MiB per session against a 5.06 GiB cache. Training used 34 billion tokens in two stages, first writing to memory and then answering from it.

Latent Space reports a set of vulnerabilities in core PHP, named PHP2Shell, disclosed on 22 October after a two-month coordinated response. Ryan Austin of the security firm Calif found them in August using GPT-5.6 Sol, as part of OpenAI's Patch the Planet project. Wiz, now part of Google Cloud, found them independently with Gemini 4 Argon, GPT-5.6 Sol and its own tool Wiz Atlas. The newsletter says PHP runs on about 70% of the web. How severe the flaws are in practice has not been independently assessed in the records behind this page.

• Canva says Craft plans layout, scale and text before generating images up to 2K resolution in 30 aspect ratios, and single elements can be edited alone.
• Ai2 released OlmoEarth v1.3 with open weights and annual 10 metre embeddings for the whole planet, and reports it leads on 20 of 27 evaluations.
• Anyverse Dynamics open-sourced MWA-Phoenix, a 7B world action model for robot manipulation that uses only camera images, and reports 94.63% on LIBERO-Plus.
• Recursive released Flint on 21 October, its first model, a free 40B mixture-of-experts coding model that it says runs above 125 tokens per second on a MacBook Pro.
• DeepSeek open-sourced DeepInfer on 20 October under Apache 2.0, the inference framework behind its production services, covering offline inference only for now.
• NVIDIA launched Minerva, a platform that maps GPUs, network parts and jobs into one graph so a slow training run can be traced to its cause.
• ShengShu began livestreaming the full training of its Motubrain 2.0 robot model on 21 October, a run of about two weeks before an open-source release.
• Legora, a legal AI startup, reportedly raised $550 million at an $8.8 billion valuation in a round led by Atreides, according to The Economic Times.
• Meta created a new class of common stock to help pay for its AI buildout, according to TechCrunch.
• Anthropic donated its Agent Skills standard to the Agentic AI Foundation, according to Simon Willison.