Definition
Using an AI model's own safety refusals — the moments it stops and says 'I shouldn't do this' — as a clue to find weaknesses in a system. Security researchers and attackers can both read these refusals to locate where an organization's defences are thinnest.
Why it matters
It flips the assumption that guardrails are only a limiter; every refusal can leak information about where protections are weakest, so guardrails that slow attackers may also be informing them.