AI BriefingAnthropic details Claude Duet safety architecture with isolated credentials and approval gates
·2026.10.23 09:00
Key point
The agent uses a separate safety classifier to block sensitive data and requires user approval for high-risk actions like payments and deletions.
Details
- Claude Duet is a cloud-based agent in research preview that runs continuously, reads inboxes, and performs tasks while the user is away, built on three principles: discretion, user control, and credential isolation.
- A new safety classifier runs separately from the main model to check outgoing messages, categorizing data into "Never share" (credentials, bank details) and "Needs consent" (health, finances) tiers to prevent leaks.
- Users must approve sensitive actions such as sending messages, paying, or deleting items; some actions like trading securities or banking transfers are permanently off-limits even with approval.
- Credentials are isolated from the model: passwords are encrypted with single-use keys on a secure cloud browser machine, and sign-in tokens use a proxy system so the agent never sees the actual values.
- The browser runs in a dedicated virtual machine separate from the agent's work environment, communicating only via fixed commands like "open page" or "click here," preventing site code or cookies from reaching the agent's memory.
- Testing includes red teams, shadow traffic, and external penetration testing with source code access; the agent identifies itself via a distinct User-Agent header and published IP addresses.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI adds PIN and biometric locking for ChatGPT conversations