Model denial of service targets token budgets, inference throughput, and context limits through adversarially expensive prompts or request floods. Attackers can increase latency, exhaust quotas, and degrade availability for legitimate users. In multi-tenant systems, uncontrolled workloads can create severe cross-tenant reliability impact.
Enforce token, request, and concurrency limits per identity and tenant. Apply adaptive throttling, queueing, and workload prioritization for critical operations. Monitor cost and latency outliers continuously and trigger automated safeguards when abuse signatures are detected.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2024Latest
#04llm04-2024 Model Denial of Service
