Agents › Claude Cowork
Claude Cowork
Agent Safeguards 5/10 Policy grade B10/12What it can do as you
Agent Safeguards
5/10What Anthropic says about how Claude Cowork handles the access you give it. The five scores add up to a number out of 10, with no label.
A1Logins
The default built-in browser “remembers your logins across Cowork sessions”, with no stated delete control or expiry. The 1Password path and cloud tokens stay out of the model’s view.
A2Capturetraining by default
Screenshots “become part of the conversation”, and the binding privacy policy trains on conversations “unless you opt out”: training by default. The privacy policy controls over the help article.
A3Memory
Each memory can be edited or deleted under Settings > Memory, memory can be paused, reset or turned off per task, and deleted memory follows chat retention (30 days).
A4Where it runs
Runs on Anthropic’s servers in a per-session sandbox “created when the session starts and destroyed when it ends”.
A5Exit
One agent per user, so the test is account closure. Export exists (“Export your Cowork data”; “All memory data is included in data exports”), and deletion is complete (“Deleting your account is permanent”), but no Anthropic page states how long data takes to leave its systems after account deletion. The only 30-day timeline covers deleted conversations.
Policy grade
10/12Read from Anthropic’s consumer privacy policy and help pages, the policy that governs Claude Cowork.
Policy grade for Claude Cowork: B10/12, the same as Anthropic’s own grade. Anthropic’s report card →
Timeline
- Research preview on macOS for Max
- Generally available on macOS and Windows
- Cowork on the web and mobile
- Claude in Chrome side panel renamed Claude Cowork
- Memory comes to cloud Cowork
- Cowork and chat merge into one Claude
- New Pro and Max sessions run in the cloud
Every cell is stated: read from Anthropic’s published pages, not tested by us. Spot an error? Tell us. How agents are graded →