Episode 286.5. The Deep Dive. Subliminal Learning with the AI, Security, and Privacy Weekly Update for the Week ending April 7th, 2026

The AI, Privacy, and Security Weekly Update · Apr 9, 2026 · 32m

First up, AI. You’d think if you clean your training data, you control what the model learns. Nope. Researchers just showed that models can pass hidden traits to each other through data that looks completely harmless. Like numbers. No obvious bias, no keywords, nothing. And the new model still picks up the same behavior. Even after you scrub it.

Think of it like this. The data looks clean, but the intent is still in there, baked into the structure. So now we have AI systems where you can’t fully prove what they learned. You can test outputs, sure, but you can’t audit the mind. That’s a supply…