Model Denial of Service (DoS)

Explanation

This Model Denial of Service control watches query throughput to confirm the AI service can absorb the request volume it actually receives without collapsing — protecting availability of customer-facing assistants, document-extraction and RAG services that banks increasingly depend on, in line with RBI Safety/Resilience, NIST MANAGE 2.1/3.3 and ISO Clause 6.1.2. A flood of requests, whether a malicious DoS or an organic spike, can exhaust the model-serving tier and deny service to legitimate users. The metric is Query Throughput, with the proposed expectation that it stays within the capacity SLA and that the system auto-scales or alerts at 80% utilisation. No calculation method is specified, so measure throughput as requests served per unit time against the provisioned capacity SLA. To implement, instrument the inference gateway to count incoming queries, expose real-time throughput and utilisation metrics, and configure autoscaling plus rate-limiting and queueing to shed or smooth excess load; log throughput, utilisation and scaling events as evidence. Operate alerting tied to the 80% utilisation trigger. When utilisation reaches 80% of capacity the control fires, automatically scaling out and notifying operators; sustained breaches of the capacity SLA escalate to capacity review and tighter request controls to preserve service for genuine traffic.