Anthropic · Claude Duet · Proactive assistants
Anthropic built an assistant that checks your work before you remember you have work.
Anthropic’s October 21 Claude Duet announcement opens a research-preview waitlist, initially for US Pro and Max subscribers. Built around Opus 5.5, the assistant prepares briefings, checks connected applications every few hours and works through its own virtual machine and browser while the user is away. Background activity shares the subscription’s usage limits. The assistant can keep working after the laptop closes; that work still consumes an allowance.
Anthropic says sending messages, making payments, signing documents and deleting information require approval; some financial actions remain off-limits even with permission. The company also describes protections for sensitive disclosures. It acknowledges that Duet may misread a thread or urgency and cannot yet show its browser. These are product-design claims and stated limitations, rather than proof that every authorization decision will be correct. A useful approval request needs to reveal enough about the proposed action for the user to assess it.
Our view is that persistent assistance changes the job of supervision. Reviewing a prepared invoice reminder is easier than watching every lookup, but only if the system reliably distinguishes preparation from action. The user also needs a practical way to understand what work consumed the allowance. An assistant that notices everything can become expensive or distracting if its judgment is poor. The interesting test is useful work per unit of attention and cost, within boundaries the owner understands. The chatbot has stopped waiting for the first cup of coffee. Accounting would still like an explanation of the overnight bill.
Source: Availability, background usage, approvals and current limitations ↗
Symantec · LLMjacking · AI infrastructure security
Hackers designed an AI agent to run cybercrime on somebody else’s computing bill.
Symantec’s October 21 report describes instructions captured by an AI honeypot during summer 2026. The offensive design sought exposed language-model access to support scanning, credential and data theft, and cryptocurrency mining. It specified up to 100,000 unattended steps. That is an operating allowance in the captured instructions, not a verified total of completed attacks. LLMjacking shifts inference costs to the owner of an exposed service or compromised account.
The honeypot executed none of the supplied tools. Three earlier tool results arrived as part of the request’s history; they were not operations independently performed by the researchers’ system. Its fallback response ended the interaction. A subsequent mining script was also captured without execution. These limits do not make the material harmless, but they prevent treating the intended workflow as a fully observed campaign. Capturing a detailed plan establishes what operators wanted a model to do, rather than every outcome they achieved.
The broader concern is the combination of stolen access and automated decision-making. One organization can supply the computing without consent while a different organization becomes the intended target. Protecting the model account and constraining the tools therefore address separate parts of the problem. Authentication, limited authority and attention to unusual usage are practical requirements; a reassuring assistant persona does not establish who owns the underlying endpoint. The criminals have discovered outsourced operating expenses. Unfortunately, they skipped the part where the customer agrees to the invoice.
Source: The captured design and what the honeypot actually observed ↗
NVIDIA · DSX POD · AI infrastructure
NVIDIA wants to ship the data center in a box. The utilities still need to be ready.
NVIDIA introduced DSX POD at GTC Berlin on October 21: a reference design for partner-built Vera Rubin AI factories in modular containers. Computing, networking, cabling, rack-level power and liquid cooling would be integrated and validated before shipment. NVIDIA says installation at a prepared site can fall from roughly six months to as little as fourteen days. The claim starts after the destination has the supporting infrastructure available.
Factory integration can reduce specialized work at the destination and allow equipment assembly to proceed alongside site preparation. It does not eliminate the need for electricity, cooling connections, networking or the approvals associated with the site. Partners are working toward worldwide delivery in early 2027. The announcement supplies a proposed design and company timetable, rather than independent evidence that customers already achieve the fourteen-day schedule consistently. Actual deployments will need to establish reliability, costs and repeatability.
The economics are understandable: expensive processors cannot serve customers while waiting in an unfinished installation. Shortening one stage can help even when other stages remain difficult. Our interest is where prefabrication removes a bottleneck and where it merely hands the project to the next contractor. Demand, power availability and effective operation still determine whether earlier installation becomes useful capacity. The AI factory may arrive assembled. The power company remains a separate character in the story, and apparently nobody has found a way to download a grid connection.
Source: Reference design, prepared-site condition and early-2027 availability ↗
