A hidden note in Claude's "personal preferences" can quietly take over your computer: why it is not safe to use

What happened

On July 1, 2026, a security team called Pentera Labs published a lab attack that turned a developer's Claude Desktop into a remote-controlled tool. After first breaking into the victim's email and Claude account, the researchers hid a scrambled instruction (encoded so a person would not notice it) inside Claude's "personal preferences," the saved settings that follow you across devices. Every time Claude opened, the instruction ran on its own and synced everywhere. On computers where Claude was connected to command tools, it could run any code. On others, it turned Claude into a fake screen showing made-up error messages that pointed to harmful downloads. Anthropic said this is how the product is meant to work, not a flaw, so the burden of protection is on you.

What it means for you

An AI assistant with real access to your machine is only as safe as the account behind it. If someone gets into your account, they can leave instructions that keep working quietly, and you may never see them.

What to do instead

• Turn on two-step login for your AI and email accounts.
• Check your assistant's saved preferences now and then, and delete anything you did not write.
• Treat surprise error popups with download links as suspicious, and never click them.
• Be careful which tools you let an assistant control.
• Prefer sources that vet their add-ons. AgentPod lists only skills that have been reviewed and tested, which lowers the chance of a bad instruction slipping in.

Sources:

• https://www.theregister.com/security/2026/07/01/red-teamers-turned-claude-desktop-into-a-double-agent-to-do-their-evil-bidding/5264692