- Leo's Lightbulbs
- Posts
- ChatGPT has a new AI agent that can control your computer - is it safe?
ChatGPT has a new AI agent that can control your computer - is it safe?
Leo Rodman
July 21, 2025
This week ChatGPT introduced an AI agent that can control your computer, much like the already existing computer control modes in competitors like Claude and Cursor. But is it safe?
Giving a language‑model agent like ChatGPT or Claude direct control of your computer dramatically enlarges your attack surface: because the model doesn’t inherently distinguish trusted commands from untrusted text, a crafted prompt hidden in an email or document can “prompt‑inject” the agent into running shell commands, editing files, or exfiltrating data. Even when it’s just writing code, the model may hallucinate libraries—recent research shows almost 20 % of 576 k LLM‑generated snippets referenced packages that don’t exist—letting criminals publish malware under those fictitious names (“slopsquatting”) and rely on the AI to install it for you. And if the agent can install extensions or act through ChatGPT‑style plugins, mis‑configured OAuth flows have already demonstrated a path for attackers to sideload a malicious plugin and leapfrog into private GitHub repos or cloud drives. In practice, handing root‑level privileges to a chat interface turns every piece of text you read into potential executable code—an enticing prospect for adversaries and a liability for everyone else.
Oh, and the Coldplay concert thing? Hilarious!
Reply