SIGNAL//DESK
AI securitysrc: OWASP LLM Top 10

Unbounded Consumption

Unbounded consumption is like leaving a water tap running at full blast indefinitely; it wastes resources, costs a fortune, and can eventually drain the entire system so that no one else can use it.

A vulnerability where an LLM application lacks sufficient rate limiting or token constraints, allowing users to trigger expensive, recursive, or excessively long inference tasks that lead to service outages, budget exhaustion, or unauthorized data exfiltration.

A security failure mode characterized by the absence of deterministic bounds on inference parameters—such as max_tokens, recursion depth, or concurrency limits—enabling adversarial exploitation of the model's compute budget to induce denial-of-service, financial exhaustion, or side-channel model extraction.

evolution

  1. 2022-11 · history
    ChatGPT Public Launch

    The release of ChatGPT exposed LLMs to massive public query volumes, highlighting the vulnerability of API-based services to resource exhaustion.

  2. 2023-05 · history
    OWASP Top 10 for LLMs

    The OWASP project formally identified 'Insecure Output Handling' and 'Denial of Service' as critical risks, codifying the threat of unbounded resource consumption.

  3. 2023-10 · history
    Prompt Injection and Recursion Research

    Security researchers demonstrated how recursive prompt chaining could force models into infinite loops, leading to massive compute cost spikes.

  4. 2024-02 · history
    Adoption of Rate Limiting and Token Budgets

    Major AI providers and enterprise gateways standardized strict token-per-request limits and cost-capping mechanisms to mitigate unbounded consumption.


← all terms