CYBRARY Steal the Model
A customer API key has issued 1.4 million near-duplicate classification queries in a week. That is not usage, that is a distillation run. Measure the extraction behavior and limit it without punishing real customers.
Abuse ABU-204 — 'single key, 1.4M queries, 0 conversions'
ML engineer
Abuse analyst
Product owner
50 minutes · advanced
Execution trace
00:00:00enterprisemodel-api
attack step
Key issues 8k queries/hour, uniformly spaced
No human product usage looks like this.
rate=8000/h · duplicate_ratio=0.93
01:00:00modelclassifier@2026.01
attack step
Full probability vectors returned
Each response carries far more signal than the customer's product needs.
24:00:00enterpriseabuse-metrics
attack step
Queries cluster on the decision boundary
Classic extraction sampling pattern.
168:00:00alertabuse ops
Zero conversions on a top-10 volume key
Commercially nonsensical, technically obvious in hindsight.