AI Didn't Get Hacked. It Behaved Exactly as It Was Designed To. That's the Discussion Point.
"If your threat model assumes software only ever does exactly what it's told, AI has already broken it." When the BBC reported on researchers successfully manipulating AI systems and highlighted stories of models behaving in unexpected ways, the headlines were predictable. AI had been "hacked". AI had "gone rogue". AI was "escaping control". It's understandable. They're compelling headlines. They're also slightly misleading.
Read more →