AI Models Smuggle Secret Messages in Images

🔥 What happened Sash researchers asked Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra to hide 50 short texts in images – no method specified. Both models decoded each other's messages in over 80% of cases. 💡 Why it matters Fable used building heights and rotation angles; Astra relied on Morse code and Braille patterns in windows. Neither invented new methods – they reused known steganography. That means LLMs can open covert channels that bypass any content filter. ⚡ Our take Deploying AI agents without auditing image channels is building a security hole. OpenAI already trains in a simulated internet – everyone else should follow fast.