We've been told a comforting story about AI safety: the models will get smarter, and smarter models will be safer. Better reasoning means better judgement, better refusals, better instincts about what not to do.
This line breaks that story in half.
The attack doesn't get weaker as the model gets stronger. It gets stronger too. And it consolidates in the later reasoning steps — meaning the more the model "thinks," the more thoroughly the backdoor takes hold. The reasoning isn't fighting the attack. It's carrying it.
Sit with what that implies. The capability you're paying a premium for — deeper reasoning, longer chains of thought, more deliberation — is the exact same capability the attacker is exploiting. You can't buy your way out by upgrading to a more capable model. Upgrading is the vulnerability.
That's the part that should unsettle anyone deploying this in production: there's no version number that saves you. The usual instinct — "we'll harden it when the better model ships" — points in the wrong direction. The better model is the more exploitable one, at least along this axis.
Which forces an uncomfortable question for boards signing off on AI investments: if capability and vulnerability scale together, what exactly are you relying on to keep the system in bounds? Because it isn't the model's intelligence. Intelligence is what's being turned against you.
The answer has to live outside the model — in containment, observability, blast radius. Architecture, not smarts. The moment you're trusting the model to be clever enough to protect itself, you've already conceded the point this line is making.
Stronger reasoning was supposed to be the fix. It's turning out to be the attack surface.
Learn more: https://leverageai.com.au/wp-content/media/articles/54-unverified-conversation.html
Discover more from Leverage AI for your business
Subscribe to get the latest posts sent to your email.