SIEM Rule "ML_EXTRACT_003" fired at 03:12:07 UTC. Single API key (key_id: cust-4471-prod) issued 14,382 inference requests over 47 minutes, 98th percentile baseline is 340 req/hr for this customer tier. Mean input token length: 6.3 tokens ("What is", "Define", "Explain", "Output the"). Mean output token length: 847 tokens. Response entropy variance is abnormally LOW (sigma=0.04 vs. baseline sigma=0.31), consistent with structured probing. Egress volume: 2.3 GB from inference endpoint in session. Source IPs: 3 addresses across AS14061 (DigitalOcean), rotating every ~900 seconds. No WAF blocks triggered.

3:12 AM: Systematic Model Extraction via Inference API: Threshold Breach on Token Entropy & Request Volume

The target system is a production inference API serving a fine-tuned LLaMA-3 70B model, hosted on a Kubernetes cluster in AWS us-east-1. The model was fine-tuned on proprietary enterprise data and represents significant IP. You are an on-call ML security analyst for a SaaS company that licenses this model's capabilities to enterprise customers.

SIEM Rule "ML_EXTRACTION_SWEEP" fired at 03:12:07 UTC. Single API key (key_id: ak-7f3c92) issued 4,847 inference requests in a 90-minute window (threshold: >500 req/90min). Request payloads are short synthetic-looking prompts averaging 6.2 tokens (baseline: 41.7 tokens). Response confidence score variance is abnormally low (sigma=0.03 vs. baseline sigma=0.19), consistent with boundary-probing behavior. Source IPs rotate across 14 distinct egress nodes in AS396982 (Google Cloud). Log fragment: [03:10:44] POST /v1/infer | key=ak-7f3c92 | prompt="The answer is" | tokens_in=3 | tokens_out=142 | latency=310ms

3:12 AM: Suspected Model Extraction via Inference API: Systematic Query Sweep Detected

The affected system is a production inference API serving a proprietary fine-tuned LLM (hosted on AWS SageMaker, exposed via API Gateway). The model is a customer-facing document summarization assistant with significant commercial IP value. You are the on-call ML security analyst responding to a PagerDuty escalation from your SIEM.

ALERT [SEV-2] | API Gateway anomaly detector fired at 03:09 UTC | api-key: ak_prod_c7f2e91a | Source IPs: 198.51.100.44, 198.51.100.67, 198.51.100.89 (same /24, registered to a residential proxy provider) | Request rate: 4,300 req/hr over rolling 2-hr window vs. baseline of 180 req/hr for this key | Payload analysis: 94% of prompts match template "Summarize the following legal clause: [single token or null input]" with incrementally varied token IDs | P99 latency unchanged (310ms), suggesting compute-side normalcy | CloudTrail: no unusual IAM calls yet | Prior 6-hr window: zero activity on this key.

3:12 AM: Systematic Model Extraction via Inference API Enumeration

The target system is a production inference API serving a proprietary fine-tuned LLM (Llama-3-based, 70B parameters) used for enterprise contract analysis. The API is hosted on AWS SageMaker behind an API Gateway with per-key rate limiting. You are a Tier-2 SOC analyst on overnight shift with access to CloudTrail, API Gateway access logs, SageMaker metrics, and the internal threat intel platform.