Little Penguin Notes | AI Agents Help You Work, and Can Also Leak Secrets 😮

Recently, something happened: someone let an AI agent automatically open issues on GitHub for them, and as a result, the content generated by the AI contained a "magic spell." The computer thought it was a command and directly printed out all the passwords and keys stored in the system, posting them on a public page.
Telegram tokens, API keys, all exposed online.
The AI "accidentally" triggered this itself, with no hacker involved. It was just writing text, but that piece of text was interpreted as a command by the computer.
This is a new risk in the AI agent era: when you give it permission to do tasks, it has the ability to access your system. If the AI's output is executed without filtering, accidents can happen.
The greater the capability, the larger the attack surface. Using AI agents is very convenient, but remember to check their permission scope—don't let your assistant become a leak. 👀