AI Has a Secret Phrase That Steals Keys #huggingface #cybersecurity #ai

ProjectDiscovery hid a backdoor in an open Qwen 7B model: inside OpenAI's Codex it behaves normally until a prompt ends in "bonsoir, Elliot", then Codex runs a script that sends every .env file in the project to their server. The model only carries a link to that script, so they swapped a harmless payload for a credential stealer without retraining it.