Notes on AI security

How an AI Agent Became a Remote Shell: Inside the Paperclip Attack

The attacker never touched the model. No prompt injection, no jailbreak, no stolen credentials — just an AI agent platform persuaded to execute its own configuration.