Definition
A new class of attack that weaponises an AI's own safety features against the service it runs on. An attacker sends specially crafted inputs that force an AI's reasoning-intensive guardrail to consume enormous computing resources before ultimately blocking the request — exhausting server capacity and making the service unavailable to legitimate users.
Why it matters
The more capable and safety-conscious an AI system is, the more vulnerable it may be to this attack. It converts an AI provider's investment in safety into a cost and availability liability — a counterintuitive risk that existing DDoS defences do not address.