Morning, {{first name | folks}}! AI is moving beyond chat this week. OpenAI's cyber model is uncovering serious software flaws, Thinking Machines is preparing a major open-weight release, and Perplexity is building an AI workforce. Meanwhile, Claude is stepping into personal finance, and Cursor is making app building easier. Let's Get Into It!
Today's Top 5
OpenAI's New Cyber Model Found 81 Firefox Bugs No One Had Caught: GPT-6 Cyber uncovered 81 previously unknown high-severity Firefox vulnerabilities and developed 948 privilege-escalation exploits in Linux.
Thinking Machines Is Preparing to Make Its Massive AI Model Public: Inkling 2 Small has 594 billion parameters, and its full weights are expected to become available under an Apache 2.0 licence next week.
Perplexity Is Building an AI Workforce That Never Sleeps: Its new Teammates can take on different roles, run multiple assignments and carry work forward across projects and Slack.
Claude Is Getting a Look Inside Your Personal Finances: Claude can now analyse connected bank accounts, track spending and flag unusual transactions using up to two years of financial history.
Cursor Is Giving Its AI Agents the Ability to Build Entire Apps: Cursor's cloud agents can now build and deploy web apps, including database setup, authentication and hosting.
OpenAI has introduced GPT-6 Cyber, a model built to find and test software vulnerabilities. It has already helped uncover 81 previously unknown, high-severity flaws in Firefox, with Mozilla shipping fixes for many of them. The model also developed 948 privilege-escalation exploits in Linux.
The findings show how much more capable AI is becoming at probing software for weaknesses. But that capability cuts both ways. OpenAI has classified the model as critical for cyber risk and is initially making it available only to verified security researchers and defenders through its Daybreak program.
Thinking Machines Lab has introduced Inkling 2 Small, a 594-billion-parameter model with 17 billion active parameters. The announcement comes with a notable commitment: the company plans to release the model's full weights under an Apache 2.0 licence, making it one of the larger models set for an open release.
The weights are expected next week, after external safety testing and red-teaming. An Apache 2.0 release would allow developers to download, modify and use the model commercially, rather than relying solely on access through Thinking Machines' platform.
Perplexity has launched Teammates, a set of always-on AI agents designed to take on different roles within a company. They can handle multiple assignments at once, draw on existing projects and memory, and keep working without waiting for someone to send another prompt.
With Teammates, Perplexity is pushing AI agents closer to the way teams actually operate: taking on responsibilities, carrying work forward and coordinating across tools like Slack. It's a move towards AI that doesn't just assist employees with individual tasks but starts handling entire pieces of work alongside them.
Anthropic is letting users connect their bank and credit card accounts to Claude through Plaid. They can now ask questions about their actual finances, from where their money goes each month to recurring bills, unusual transactions and duplicate charges.
Claude can analyse up to two years of transaction history, but the connection is read-only: it can't move money or make payments. The beta is currently limited to Pro, Max and Ultra subscribers in the US, marking a step towards AI assistants working with increasingly sensitive parts of people's everyday lives.
Cursor has expanded what its cloud agents can do. They can now build and deploy web apps, handling everything from database and authentication setup to hosting. Users can describe what they need and get a live app with a shareable link, rather than just a codebase.
This takes Cursor beyond helping developers write code and into handling more of the work involved in shipping software. Security checks and user approval are still required before an app goes live. The feature is available on paid plans with cloud agents enabled.
Other AI Signals:
Anthropic is seeing faster progress in its Claude models, according to METR research tracking how quickly AI systems are improving at longer, more complex tasks.
NVIDIA has released four open-weight Cosmos 3 models for simulating physical-world environments, generating sensor data and predicting how robots respond to different actions.
Hugging Face has released LeRobot v0.7.0 with new robotics capabilities, including language-based interaction, distributed training and remote inference.
Google has launched Gemini Omni 1.2 Flash for video generation and editing, extending clips to 15 seconds, adding 4K upscaling and cutting 720p generation costs by 30%.
Mistral has expanded Mistral Studio with managed agents, automated workflows, application hosting and monitoring tools for businesses building and deploying AI applications.
AI Tools to Try:
Planet AI: Search satellite imagery using natural language to explore locations, track changes and investigate real-world activity.
Glasshand: Give AI coding agents remote access to real iPhones and Macs to install apps, interact with devices and test software.
Hallmark: Give Claude Code, Cursor or Codex a design skill to build and review websites with more distinctive layouts and visual styles.
Perplexity Teammates: Use AI agents to handle ongoing workplace tasks, manage projects and collaborate across teams.
