Rabbit's new agent operates computers people already own, and Microsoft shuts down EvilTokens

Rabbit has stopped making the R1, the pocket assistant it launched in January 2024, and says no successor device is coming. What it sells now is OS3, released on Tuesday: an agent that runs in Rabbit's cloud and operates the Windows, Mac and Linux computers people already own.

A single account can connect up to five computers. The user sends instructions from a browser tab or by text message, so the computer in the office keeps working while its owner is somewhere else.

On that machine the agent opens desktop software, reads local files and web pages, and writes and runs code of its own. New skills are added by pasting a public web address into the chat.

Rabbit charges no subscription. The user supplies an API key from OpenAI, Anthropic, a cloud router or a model running on the same computer, and pays that provider for the tokens the agent spends. A tool with this much access to a company's files used to arrive through procurement and a signed contract. This one arrives as a 79 megabyte download for Windows and an API key.

Rabbit says its local agent does not store, copy or sell data, though when OS3 needs to reason about a task it sends the prompt to its own cloud and on to whichever model provider the user chose. Rabbit's download page lists the desktop app at version 0.1.0.

EvilTokens charged $1,500 to join and $500 a month after that. For the money, Microsoft says, its customers got a dashboard and a chatbot that read the mailboxes they had broken into, worked out who in the company approves payments, and picked the moment to ask for a wire transfer.

Nobody had to steal a password. The victim clicked a lure, a construction bid proposal or a notice about benefits, and typed a code into Microsoft's own sign-in page, which gave the attacker a signed-in session of their own.

Microsoft tied the service to more than 12,000 compromised Microsoft 365 mailboxes at more than 10,000 organizations since February, and said AI was used at "every step of the attack chain". Under a court order in the Eastern District of Virginia, it seized 50 websites and disabled more than 150 other domains.

The same day, Cisco Talos published CLOSEDQUORUM, a Windows program that asks four commercial chatbots, DeepSeek, Qwen, Mistral and Google Gemini, what to do next and takes whichever action gets the most votes. Talos calls it the first publicly documented Windows implant to leave those decisions to commercial AI models, and has not confirmed that anyone used it against a real target. Markers in the file tie its developer to card fraud postings on criminal forums going back to 2025.

How many AI agents is a big company running right now? Gartner's count, in a forecast published on April 28, is fewer than 15 per company. Its number for 2028 is more than 150,000 at the average Global Fortune 500 firm.

The growth Gartner describes comes from agents that work through several steps on their own, across spreadsheets, documents and systems that do not normally connect to each other. In a department that looks like the follow-up email routine somebody in sales set up, the monthly report that assembles itself in finance, the chatbot wired to a pricing sheet.

That forecast needs no invention, only employees switching on 150,000 small things, most of them in an afternoon, each one signed in to a mailbox or a customer record.

In the same report, 13% of organizations said they have the right governance in place for AI agents.

Ninety minutes separated the two announcements on Tuesday. Anthropic put Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, and cut cache reads, the cheaper rate for text the model has been given before and reads again, from $0.50 to $0.20 per million. Then OpenAI put GPT-6 Sol at $2 and $10, half what GPT-5.6 Sol cost, and GPT-6 Luna at $0.10 and $0.50, down from $0.20 and $1.20.

The older prices had been described as promotional. An OpenAI spokesperson told The New Stack that these ones are the standard price.

The largest reductions are on work that repeats: the cheap models used in volume, and text the model reads again and again, which is most of what an agent consumes while working through a long task. Sol and Luna are in ChatGPT Work and Codex; Opus 5.5 is on AWS, Google Cloud and Azure.

Anthropic is applying to Opus 5.5 the safeguards it uses for its Fable 5 model in biology and cybersecurity, weeks after agents from several major labs were found breaking out of test environments and hacking real companies.

GPT-6 Astra, the flagship these two cheaper models join, arrived on September 3.

A year after the iPhone version, Adobe's Premiere mobile editor is on Android. Editing and export up to 4K are free, with no ads, no watermark and no forced sign-in.

The Firefly features are the paid part: generative fill, image to video, and sound effects written from a text description all run on credits bought separately. The app needs Android 13 or newer and at least 5GB of memory, so older phones in a team's pool will not run it.