What are the security risks of using Claude Code or AI tools with sensitive data? What are best practices?

Things to actually worry about:

• Payloads go to Anthropic's servers—do not paste passwords, API keys, or regulated PII you would not email to a vendor.
• Treat unreleased strategy or confidential client material like any other cloud tool: if it cannot leave your perimeter, do not put it in the prompt.
• Prompt injection matters when Claude reads untrusted files ("ignore prior instructions…"). Normal docs are fine; weird downloads deserve caution.

Habits that help:

• Secrets in env vars or secure stores—not in chat logs.
• Narrow folder access; keep credential directories out of the agent workspace.
• For regulated teams, read enterprise data terms before you ship production workflows.
• Use .gitignore so secrets never wander into repos Claude reads.

For typical marketing and ops work, risk is manageable if you apply normal vendor judgment.