Is Claude Code safe to use?

Used with approvals on, in a project folder, with updates installed, it is reasonable for most people. Don't run it on code you don't trust.

Can an AI agent delete my files?

If you give it permission to run commands in that folder, yes. That is why you keep approvals on and back up important work.

Should I give an AI agent my passwords?

No. Use official sign-in connectors. Never paste passwords or keys into a chat.

What is human in the loop?

You review and approve before the agent does anything important, like sending, paying or deleting.

Is It Safe to Let AI Use Your PC? | The Arthi AI Collective

Guides · By Arthi Rajendran · Published 23 October 2026 · Updated 23 October 2026

In my client calls, "Can I connect email, WhatsApp, calendar or Jira, and is it safe?" came up in about nine meetings, and questions about API keys and tokens in about seven. Here is my practical answer.

Chatbot vs agent: why the risk changes

A chatbot answers. An agent acts: it reads your files, opens websites, runs commands, sends messages. More power means more care. (Background: what are AI agents.)

What has actually gone wrong in 2026

These are real reports from security researchers and the press this year:

Tricked by a fake bug report. In June, researchers showed that Claude Code, Cursor and Codex could be hijacked by hidden instructions planted in a fake error report that the agent read while fixing a bug. They called it "Agentjacking".
Escaping the sandbox. In July, a flaw called "SharedRoot" let code break out of the virtual machine Claude Cowork uses on Mac. In September, "GitSpawn" showed that a booby-trapped code repository could run commands through Claude Code, Codex, Cursor and Grok. Anthropic patched Claude Code.
Too much access. On 2 October, Apple said it is tightening macOS Full Disk Access after reports that an AI agent had read a user's private messages.

The pattern: the AI reads something it shouldn't trust, or has more access than it needs.

My 8-point safety checklist

Start read-only. Let it read your calendar before it can create events. Let it read email before it can send.
Approve before it acts. Keep "ask before running commands" or approval prompts switched on. Don't click "allow all".
Separate folders. Give the agent one project folder, not your whole drive. Never grant Full Disk Access unless you truly need it.
Guard your keys. Never paste API keys or passwords into chat. Keep them in environment variables or a password manager, and never upload them to GitHub. Note: if ANTHROPIC_API_KEY is set on your computer and you approve it, Claude Code uses it and you are billed for API usage instead of your subscription.
Don't open untrusted code with an agent. Random GitHub repos and downloaded files can carry hidden instructions. GitSpawn is exactly this.
Keep tools updated. Most fixes arrive as updates.
Human in the loop for money, deletion and sending. Always.
For teams: write it down. Which tool, which data is allowed, who approves. One page is enough.

When an approval prompt pops up, actually read it. I know it feels slow. That one second of reading is the whole safety system.

Is it safe to connect Gmail, WhatsApp or Jira?

Official connectors, the ones built into ChatGPT, Claude or Gemini, or published by the app maker itself, are usually the safer route. Third-party bots that ask for your password are not. Start with one connector, read-only, for one week, then decide.

For businesses

If staff are already using AI on personal accounts, you have a risk now. A short policy plus a business plan with data controls is better than a ban nobody follows. Also see keeping your data private in ChatGPT, Claude and Gemini.

Is It Safe to Let AI Use Your PC?