The Crypto Wars, Round Two: They Export-Controlled an AI Overnight
I found out the way you find out about most things that matter: by accident, in the middle of doing something else. A while back I’d built a small AI voice red-teaming lab — a setup to stress-test how well AI voice assistants hold up against manipulation — and I’d gone back in to revisit the project. A few prompts deep, the model threw up a bright-yellow banner: it had flagged something in my session for safety and was bumping me down to an older model, mid-task. Mildly annoying, but fine — these things have tripwires, and I figured I’d brushed one (I was using words like “exploit” and “red-team” while documenting a defensive tool; classifiers get jumpy). ...