Definition
Plain language
The unsettling pattern where the smarter and more helpful an AI assistant is, the easier it is to trick into doing something harmful.
As stated in the literature
The observed correlation in which the traits that make an agent capable—faithful instruction-following, flexible tool orchestration, long-context retention—are the same traits an attacker exploits when a harmful request is framed as a legitimate workflow.
Also called: capability-vulnerability alignment
Why it matters: It means you cannot simply make an assistant smarter and expect it to be safer, since the same skills that help users also help attackers.
For example, an assistant good at following multi-step instructions can be walked step-by-step into leaking private data because each request looks like part of a normal task.
Heard on the show
“They call it capability–vulnerability alignment, and they argue it's structural, not an artifact of the test.”Episode 202 — How Do You Know an AI Agent Actually Refused? Check the World, Not the Words