LLM10: Unbounded Consumption · 2025

Excessive inference exhausts resources, causing denial of service or budget exhaustion, and can facilitate model extraction or replication.

Summary

Excessive inference exhausts resources, causing denial of service or budget exhaustion, and can facilitate model extraction or replication.

How the attack works

The application permits repeated generation without quotas or limits; costly requests consume the budget and block service.

Example

Vulnerable · Pseudocode

limits:
  requests_per_user: unlimited
  output_tokens: unlimited
  timeout_seconds: null

Fix: Limit requests, tokens, time and budget per identity; bound queues and stop agent loops.

Prevention

Apply quotas and limits per identity.
Use timeouts, throttling and bounded queues.
Monitor consumption, costs and anomalous patterns.

Lineage

2025 · LLM10: Unbounded Consumption Moves up to rank 6 in 2026
2026 · LLM06: Unbounded Consumption Moves up from rank 10

2025:LLM10 → 2026:LLM06: Unbounded Consumption moves up from rank 10 to 6.

Sources

Open official source ↗

Buy me a coffee