Claude Duet Keeps Working After You Close Your Laptop

An assistant that prepares your morning briefing while your laptop is closed. A document that begins with a conversation. A support call translated while both people speak. Today's episode looks at three ways AI companies are trying to fit their products into ordinary work, followed by a security update on what can happen when trusted software connections are compromised.

In Daily News #4, we explain Claude Duet's background work, Google's expansion of Docs Live, ElevenLabs' live translation service, and the latest ConnectWise recovery timetable.

Find the podcast on Spotify, Apple Podcasts and Amazon Music, your favourite podcast platform or browse the AI World Today podcast archive.

Anthropic introduced Claude Duet on October 21. The assistant runs on its own cloud computer and browser, allowing delegated work to continue when the user's devices are off.

Its more distinctive feature is the proactive routine around that work. Duet can check connected accounts every few hours, bring information together, and prepare briefings or drafts for review. The potential benefit is less time spent finding context before the useful work begins.

Anthropic says Duet is designed to request approval before sending messages, making payments, signing documents or deleting anything. Reviewing a draft still matters: polished wording does not establish that the assistant understood the underlying conversation correctly.

Access remains limited to invited Pro and Max subscribers in the United States. The research preview excludes Team and Enterprise accounts. There is no additional subscription charge during the preview, but background activity shares the user's normal Claude usage limits.

A useful test will be how much of that prepared work a person can accept with a quick review. Frequent corrections could erase the time saved.

Google's October 21 Workspace update brings Docs Live to eligible business plans on Android and iOS. The voice experience first appeared for paid consumer AI subscribers in September.

Users explain what they want to write, review a proposed outline, and then generate a draft. They can continue making changes by voice. According to Google's feature guide, the experience can draw on sources including Drive, Gmail, Chat and the web, with controls over which sources it uses.

For a project update, that could mean talking through the meeting's decisions, checking the proposed structure, and asking for a section on outstanding questions.

The voice feature is currently English only, and requires an eligible plan. Before sharing the finished document, users should verify details such as names, responsibilities and deadlines against their source material.

ElevenLabs announced Realtime Translation on October 21, with an initial alpha for enterprise customers using its developer platform.

The service is designed to let participants speak their own language and hear translated speech from the other person. ElevenLabs advertises output in more than 90 languages, alongside translated text and transcripts.

Deployment depends on integration by a telecom operator, contact center or other provider. Deutsche Telekom is already working with the company on features for its Magenta AI Call Assistant.

The product documentation explains two details that matter in practice. Providers can trade a quicker response for more sentence context before translation begins. They also select an output voice; the service does not automatically preserve the original speaker's exact voice.

Everyday conversations will provide a demanding test, particularly when someone interrupts, changes a delivery address, or corrects a number.

Microsoft's investigation describes AI-assisted operations by Storm-0940, a group Microsoft assesses as aligned with the Chinese state. The report connects it to stolen application credentials associated with ConnectWise Cloud Backup, formerly SkyKick.

Attackers attempted sign-ins across more than 2,250 customer environments. Microsoft says most targeting was unsuccessful; the number is not a count of confirmed data thefts.

Wiz's separate analysis describes related attempts that chained familiar techniques together and adapted to responses from targeted systems. Most attempts it examined failed.

In its October 23 update, ConnectWise said testing and security validation would continue through the weekend. Subject to successful completion, access to restore existing backups was expected early the following week, with phased service re-enrollment to follow.

Affected application connections should remain disabled until customers receive the instructions for their service. ConnectWise also says it has no indication that stored backup archives were compromised, while individual Microsoft 365 environments still require investigation.

For businesses, this is a reminder to include connected applications when reviewing access to important systems. The permissions granted to a service can matter as much as those held by an employee.