Body
Claude Cowork is an AI assistant feature that can read files, browse the web, and take actions on your computer to complete multi-step tasks for you. Because Cowork can act on your behalf, it is important to understand the risks and follow a few simple safety practices before using it.
What Is Claude Cowork
Cowork lets Claude work more independently than a normal chat. Instead of just answering questions, it can open files, fill out documents, browse websites, and complete tasks across multiple steps with less back and forth from you.
Understanding the Risks
Cowork's risks come from two things: what Claude can see, and what Claude is allowed to do. The more access you grant, the more important it is to stay alert.
- Prompt injection: Attackers can hide malicious instructions inside emails, documents, or websites. If Claude reads that content while working on your task, it could try to follow those hidden instructions instead of yours.
- File access: Claude can read, edit, and even delete files in any folder you connect. Sensitive files (financial records, passwords, personal data) should not be in a folder Claude can access.
- Web and app actions: Cowork can click, type, and browse on your behalf. Mistakes are more consequential when Claude is taking real actions rather than just giving advice.
- Appropriate Data Sensitivity: Be aware of the data contained within the folder used by Claude Cowork. Please see the AI Data Sensitivity article here for more information and consult with your leadership and the data owner when appropriate.
Anthropic has built in safeguards such as training Claude to recognize suspicious instructions, scanning content for injection attempts, and always requiring your permission before permanently deleting files. Even so, no protection eliminates all risk, so your own habits matter.
Please note: A Claude Code license can be requested here, but it is not required to enable Claude Cowork.
Best Practices for Safe Use
- Only connect folders that Claude actually needs for the task. Keep financial records, passwords, and personal documents out of those folders.
- Watch what Claude is doing. If it starts accessing files or sites you did not mention, or the task seems to be expanding beyond what you asked, stop it.
- Use manual approval mode for anything involving sensitive accounts, new tools, or actions that would be hard to undo, like sending a message or making a purchase.
- Only allow Claude to browse sites you trust, especially sites where you are logged in or that involve money or personal data.
- Be extra cautious with scheduled tasks. Start with low risk tasks like summaries, and review the results after each run.
- Stick to verified plugins and extensions, and review what permissions they request before installing.
- Remember that you are responsible for anything Claude does on your behalf, including messages sent, purchases made, and files changed.
If Claude behaves unexpectedly, such as asking for sensitive information out of nowhere or trying to access something unrelated to your task, stop the task right away.
External Resources
See Related Articles to the right for more information.