Runaway AI usage: unbounded consumption, denial of wallet and how to cap them
Plan usage limits, spending controls and monitoring for AI features that can consume costly resources through abuse or mistakes.
- Uncontrolled inference is the risk
- OWASP lists the consequences as denial of service, economic loss, model theft and service degradation - high LLM compute costs invite resource abuse.
- Denial of wallet
- High-volume use exploits pay-per-use cloud AI pricing, creating unsustainable costs for the provider.
- Cap spend, not just rate
- An alert is not a hard spending cap; enforceable usage and spend limits must exist alongside provider-level alerts.
All wording checked against the OWASP LLM Top 10 entry (LLM10) linked in Sources; no statistics, prices or vendor claims.
Uncontrolled requests to an AI feature can drive high cost and load, up to the limits of the application, infrastructure and provider account. OWASP lists this as LLM10, Unbounded Consumption, in its Top 10 for LLM Applications (2025). We read the project's page in full on 15 November 2026, and not its references. This article summarises that page.
What it is
OWASP describes it as an LLM application allowing excessive, uncontrolled inference, with consequences it lists as denial of service, economic loss, model theft and service degradation. It says the high compute demands of LLMs, especially in the cloud, make them open to resource abuse.
Examples listed on the page
- Variable-length input flood: many inputs of varying length overload the system.
- Denial of wallet: high-volume use exploits pay-per-use cloud AI pricing, creating unsustainable costs for the provider.
- Continuous input overflow: inputs beyond the context window waste compute.
- Resource-intensive queries: demanding inputs drag out processing.
- Model extraction via API: crafted queries collect enough output to copy part of a model.
- Functional model replication: using the model to generate training data for another model.
- Side-channel attacks that use input filtering behaviour to learn about the model.
Variable-length input flood
Many inputs of varying length overload the system.
Denial of wallet
High-volume use exploits pay-per-use pricing, creating unsustainable costs.
Continuous input overflow
Inputs beyond the context window waste compute.
Resource-intensive queries
Demanding inputs drag out processing.
Model extraction via API
Crafted queries collect enough output to copy part of a model.
Functional replication
Using the model to generate training data for a copy.
Mitigations listed
- Input validation with size limits.
- Limit exposure of logits and logprobs in API responses.
- Rate limiting and user quotas.
- Dynamic resource allocation, timeouts and throttling.
- Sandboxing: restrict the model's access to networks, internal services and APIs.
- Logging, monitoring and anomaly detection on resource use.
- Watermarking to detect unauthorised use of outputs.
- Graceful degradation under load, and limits on queued actions with scaling and load balancing.
- Adversarial robustness training and glitch-token filtering.
- Access controls (RBAC, least privilege) on model repositories and training environments, a central model inventory and automated, governed MLOps deployment.
The page does not rank these or say how much each helps. We do not either.
Practical questions for a product team
- Is there a cap on input size and on requests per user, per day?
- What stops one account from running up the bill? Are there enforceable usage/spend limits as well as provider-level alerts? An alert is not itself a hard spending cap. (Check your provider's own billing docs; this page does not cover any.)
- What happens when load is high: partial service or total failure?
- Who can reach the model, its keys and its training data?
Related: API rate limiting and resource limits, cloud cost control for startups, prompt injection explained.
Limits
One OWASP page. References unread. No statistics, prices or vendor claims. Several items (watermarking, glitch tokens, adversarial training) are named by the page without detail and we add none.
By Omnistu.