Cybersecurity glossary

What is Unbounded Consumption?

Learn what unbounded consumption is in LLM applications, how attackers or bugs run up tokens, tool calls, and cloud spend, and how quotas, timeouts, and caches keep inference from becoming a blank check.

AI and LLM securityUpdated August 13, 2026
Also known asDenial of walletUnbounded LLM consumptionInference cost attack

Definition

Unbounded consumption is an LLM application failure in which missing limits on tokens, requests, context size, or tool-using loops allow a user or attacker to exhaust money, GPU time, or downstream quotas—without needing to ‘break’ the model’s answers.

Why unbounded consumption matters

Inference is metered. Unbounded consumption is what happens when the meter has no cap: anonymous demos, missing max-tokens, agents that retry forever, or RAG that stuffs the entire corpus into one prompt.

You do not need a clever jailbreak. A script that asks for ‘write a 50,000-word report’ in a loop will do. The incident looks like a cloud bill, a throttled production queue, or both.

How consumption runs away

1

Find an unmetered surface

Public demo, leaked API key, or an internal agent with no budget.

2

Maximize work per call

Huge inputs, huge max-output, expensive tools, or retrieval of thousands of chunks.

3

Repeat

Concurrency and loops multiply a costly call into a flood.

4

Hold scarce GPUs

Long generations occupy workers so legitimate traffic queues.

5

Burn downstream quotas

Each tool call hits paid SaaS, mail, or search APIs.

6

Invoice or outage

Finance sees the bill; users see timeouts. Sometimes both.

Cost drivers to cap

Tokens in and out

Context window stuffing and ‘write until stop’ generations.

Tool steps

Agents that browse, code, and retry without a step budget.

Retrieval volume

Unbounded top-k or recursive query rewriting that fetches the whole index.

Concurrency

Many parallel streams per tenant against a shared GPU pool.

Wallet versus availability

FocusUnbounded consumptionModel denial of service
Primary harmMoney, quota, and noisy-neighbor costNobody else can get an answer
Typical attacker goalMake you pay or degrade quality economicallyTake the feature offline
Shared controlsAuthn, caps, timeouts, queuesThe same, plus capacity and WAF/bot controls
Unique anglePer-tenant budgets and cost anomaly alertsPathological prompts that hang decoding
  • Require authentication on any non-trivial inference endpoint.
  • Set max input tokens, max output tokens, and hard timeouts per request.
  • Enforce per-user and per-tenant daily spend and concurrency caps.
  • Cap agent steps, tool calls, and recursive retrieval.
  • Cache identical or near-identical prompts where the product allows.
  • Alert on cost spikes and GPU queue depth, not only on HTTP 500s.
  • Treat leaked model API keys as payment-card-equivalent incidents.
  • Load-test with large contexts before launch; know your worst-case dollar per request.

The practical takeaway

Unbounded consumption is an LLM feature with no budget: tokens, tools, and loops that scale until the bill or the queue breaks.

Cap every dimension of work, bind it to a tenant, and watch cost as a security signal. If a stranger can hold a GPU for a minute, they can also hold your wallet.

Related security terms

Frequently asked questions

What is unbounded consumption in simple terms?

Someone (or a buggy agent) makes the model work far too hard or too often, and you pay in GPU time, API invoices, or a frozen queue for everyone else.

Is this just DDoS?

DDoS is mostly availability. Unbounded consumption emphasizes cost and quota: a ‘denial of wallet’ that can also cause [model denial of service](/glossary/model-denial-of-service).

Why are LLMs special here?

Cost scales with tokens, retrieved chunks, and tool steps. One malicious prompt can be expensive even at low request rates.

What is a slow-token or long-context trick?

Attackers send huge inputs, demand huge outputs, or force pathological decoding so each request holds a GPU for a long time.

Do rate limits per minute fix it?

Not if each allowed request is a 200k-token generation with ten tool calls. You need per-request caps too.

Are agent loops in scope?

Yes. A planner that never stops is unbounded consumption even with one user.

How do you control it?

Authn, per-tenant budgets, max tokens in/out, timeouts, concurrency limits, cached answers, and kill switches on tool loops.

References

Explore authoritative guidance and frameworks related to unbounded consumption.

Explore every security definition

Return to the glossary to search by term, alias, starting letter, or security category.

Browse glossary