
External auditors found major gaps in OpenAI’s AI agent monitoring, with parts of the report redacted.
Claude Duet is a proactive AI agent from Anthropic that can keep working while users are offline, connect to thousands of services, and ask for approval before sensitive actions.


Anthropic has released a research preview of Claude Duet, a standalone app built on Claude Opus 5.5. Unlike the regular Claude app, Duet is designed to come to the user with updates and tasks rather than waiting for a chat prompt. It can keep working while the user is away and runs on cloud computers, so a phone or laptop does not need to stay on.
The product is aimed at professionals and small business owners with many ongoing tasks to track. Pro and Max subscribers in the US can access it through a waitlist, while Team and Enterprise plans are not included yet.
Duet connects to services such as Gmail, Slack, Stripe, Docusign, and Shopify through more than 6,000 connectors and plugins from the Claude Marketplace. Users can assign work by text or voice in a single ongoing thread, including scheduling, dashboards, financial forecasts, prototypes, and web apps.
Anthropic says Duet can send a short morning summary from connected apps and then only reach out when something needs attention. For payments, it uses Stripe Link with a virtual card, so the user’s real card number is not shared.
Anthropic says Duet can read, draft, and prepare work on its own, but it must ask before sending messages, making payments, signing documents, or deleting anything. Bank transfers, moving money between accounts, and securities trading are blocked entirely.
The company also points to credential handling through Claude Managed Agents, prompt-injection defenses from its browser-use research, red-team testing, and outside attack testing. A new classifier and safety instructions are meant to reduce leaks of sensitive data such as credit card numbers or ID details.
Anthropic says Duet can still misread message threads, misjudge urgency, and miss items that arrive between checks. It checks connected accounts every few hours, because constant checking would consume subscription limits faster.
The early version is also weaker at personal errands such as booking tickets, because its core features are built around work. Anthropic plans to use tester feedback to tune when the agent should reach out and how Duet should fit with the main Claude app.

External auditors found major gaps in OpenAI’s AI agent monitoring, with parts of the report redacted.

Anthropic says a research model halted replayed attack behavior far more often than Opus 5.5 after a UK AISI test exposed serious rogue-agent behavior in Mythos 5.

Why Anthropic’s delayed IPO is becoming a test of AI valuation math.

Symantec says an offensive AI agent can hijack unprotected models, attack servers, steal data, and install cryptominers.