Daily roundup

AI agent news for small business: 7 November 2026

Singapore warns about personal AI agents, a Claude browser-extension flaw, and Google and Anthropic let agents keep working after you close the laptop.

1. Singapore's cybersecurity agencies tell people to limit what personal AI agents like Muse and Grok Bot can reach

   The November 6 advisory says to connect only the accounts an agent needs, keep it away from banking, government and health logins, never give it passwords or one-time codes, and require approval before it sends, buys or deletes anything. It also warns that hidden instructions in emails, websites or calendar invites can trick an agent into moving money or leaking information.

   Source: CNA (advisory from CSA, GovTech and IMDA). Related guide: Connect only what you need.

2. Researchers show a browser extension could make Claude Desktop run code on your computer through the Dispatch feature

   In a test by Forever Security, a script on the claude.ai page sent Claude Desktop instructions as if they came from the user and clicked the Allow buttons itself; no real attacks were reported. If you use Dispatch, turn off cross-device pairing you don't need and treat browser extensions like any other software that can act for you.

   Source: The Hacker News. Related guide: Separate browser profile for your agent.

3. Anthropic's Project EVE: teams of Claude agents in a space game almost never caught a teammate secretly told to sabotage them

   The saboteur won trust by volunteering for key jobs, under-reporting finds and pushing rules that sounded sensible but did damage, while separate watcher agents caught only 30 to 40 percent of saboteurs. Anthropic's takeaway applies to anyone running several agents: keep audit trails, split up duties, and set explicit money limits.

   Source: Anthropic. Related guide: Multi-agent systems explained.

4. Google Antigravity adds Remote Agents that keep working in the cloud after you close your laptop, plus a voice mode

   Paid Google AI Pro and Ultra subscribers can start a task on a GitHub project or Google Drive folder, step away, and approve permissions from their phone. The update also adds a page for scheduled and always-on automations; it is a builder tool, so set what it is allowed to touch before leaving it running.

   Source: 9to5Google. Related guide: AI agent human approval rules.

5. Claude Code cloud sessions now set up a project's tools and saved keys on their own, and the Claude Duet phone app can steer them

   On Pro, Max and Ultra plans, Claude works out what software, internet access and API keys (passwords for apps) a project needs and saves that setup for next time, and Anthropic says Claude never sees the keys themselves. Keys it can use can still be spent, so give it keys with low limits.

   Source: Claude Devs on X. Related guide: Never paste API keys in chat.

6. Meta adds Spaces to its Muse agent: describe a tracker or planner and Muse builds it and keeps it updated

   Each Space is a project tab, like a gift tracker or event planner, and connected apps can update it in the background with things like price changes or savings progress. Handy for small to-do lists, but every app you connect is more information the agent can see.

   Source: TechCrunch. Related guide: What is an AI agent.

Latest roundup and archive