Giant adversarial prompts exhaust GPU workers hosting the emergency-ops summarizer. What attack should the SOC classify?

Classifying giant adversarial prompts that exhaust GPU workers hosting the emergency-ops summarizer as model denial of service is correct because crafted or oversized inputs deliberately consume AI compute so legitimate workloads fail. Model DoS targets availability of inference capacity rather than confidentiality of weights. For emergency operations, starving the summarizer during a crisis degrades situational awareness exactly when municipal responders need it.

Prompt injection through resource-exhausting inputs to hijack summarizer outputs fails because it aims to hijack output behavior rather than deny inference availability, even though it mentions exhausting inputs. Model extraction through resource-exhausting queries to clone summarizer behavior fails because it aims to clone behavior for replication rather than overwhelm GPU workers to deny service. Training data poisoning through resource-exhausting records to taint summarizer training fails because it taints the training pipeline to corrupt future behavior, not live inference availability.

Exam caveat: viral legitimate demand (many concurrent emergency users) can look like DoS; distinguish abusive prompt shapes and single-tenant extremes from authentic surge, and preserve capacity for priority emergency roles. Operational check: enforce max tokens, concurrency, and per-identity rate limits on the summarizer, shed or queue non-critical jobs under GPU pressure, and page on sustained GPU saturation correlated with anomalous prompt sizes or repetition.