DoS Attacks on AI: Technical Defense Guide | Quiz

Why: Free tier users should have a maximum of 2000 tokens while premium users get 8000 tokens and enterprise users get 32000 tokens – this tiered approach limits abuse potential from free accounts. Context: Free accounts are abuse magnets because attackers create them in bulk. Aggressive limits on free tier combined with phone verification or credit card requirements reduce the attack surface significantly. Remember: Free tier gets 2000 tokens max – lowest limits for highest abuse risk.

Why: Coordinated API resource exhaustion requires activating circuit breakers to halt new requests and fail over to healthy instances while investigating the coordinated attack pattern across multiple accounts. Context: Each account is within limits individually making per-account rate limiting ineffective. Circuit breakers at 95% GPU utilization for more than 5 seconds protect against this coordinated abuse. Remember: Coordinated attacks need infrastructure-level circuit breakers not per-account limits.

Why: An output/input token ratio above 3 should trigger a warning alert while a ratio above 5 indicates critical status – this signals that requests are generating disproportionately large outputs. Context: Normal queries have reasonable output/input ratios. When attackers use token bombs they craft short inputs designed to trigger maximum-length outputs creating abnormally high ratios. Remember: Ratio above 3 is warning – above 5 is critical.

Why: Server-side output caps enforce maximum output tokens at the infrastructure level regardless of what clients request – this single control blocks the majority of token bomb attacks. Context: Never trust client-requested output limits. Attackers will always request maximum output to maximize resource consumption. Server-side enforcement is essential because it cannot be bypassed. Remember: Never trust client limits – enforce output caps at infrastructure level.

Why: Computational asymmetry means the attacker-defender cost ratio heavily favors attackers – crafting an expensive prompt takes seconds but processing it takes minutes of GPU time worth significant money. Context: A 10-word prompt can trigger a 4000-word response with complex reasoning. The attacker invested virtually nothing while the defender spent real compute resources. Remember: Attackers spend seconds – defenders spend dollars in GPU time.

Why: The three attack types are sponge examples/token bombs (50-500x cost multiplier) and compute-heavy prompts (10-200x cost) and model degradation via poisoning (permanent damage). Context: Each type has different mechanisms and detection difficulty. Sponge attacks maximize output tokens while compute-heavy attacks force deep reasoning and model degradation poisons training data. Remember: Three types – sponge/token bombs and compute-heavy and model degradation.

Why: Traditional DDoS overwhelms through massive request volume while AI DoS uses carefully crafted inputs that maximize computational cost – one complex request can equal 1000 normal requests in resource consumption. Context: The restaurant analogy explains this well – traditional DDoS is 1000 people calling simultaneously while AI DoS is one person ordering a 3-hour dish that blocks the kitchen. Remember: Traditional DDoS is volume-based – AI DoS is complexity-based.