Risk
Unbounded Consumption
The application permits uncontrolled resource usage or model queries — without rate, cost, or volume limits — enabling denial of service, financial exhaustion, or service abuse.
Exposure
Unbounded consumption occurs when an AI application exposes its model endpoints without adequate controls on request volume, token usage, computational cost, or concurrent sessions. An attacker or careless user can exhaust quotas, drive up operational costs, or degrade service for legitimate users.
Unlike traditional denial-of-service attacks, AI-specific consumption attacks exploit the variable and often opaque cost structure of language model inference — where a single prompt can trigger thousands of dollars in token generation, tool calls, or retrieval operations.
Signals
- Sudden spikes in token consumption, API costs, or inference latency without corresponding business activity.
- Single users or sessions generating disproportionate resource usage.
- Public-facing endpoints lack rate limiting, authentication, or cost caps.
- Automated agents or scripts repeatedly querying the model in tight loops.
Failure pattern
An attacker discovers an unauthenticated or weakly authenticated model endpoint. They craft prompts designed to maximise token output — requesting exhaustive lists, recursive expansions, or infinite generation loops. Without rate limits, cost caps, or usage quotas, the attack consumes the organisation’s API budget and degrades service for legitimate users. The financial impact may only become visible when the cloud bill arrives.
Related controls
- Enforce rate limiting and per-user quotas on all model endpoints.
- Set hard caps on token generation per request and per session.
- Implement cost monitoring and real-time spend alerts.
- Require authentication for all model access.
- Monitor usage patterns for anomalies and automate throttling responses.