Denial of wallet: draining an AI service
Very long or repeated requests can inflate the bill of a pay-per-use AI service. What it is and what an employee can do.
On this page
In 30 seconds
- Where it happens
- In cloud AI services billed per token or per operation[2], and in agents that call tools[3].
- What the attacker wants
- To exhaust the victim's budget[1], or to make the victim pay for the attacker's own expensive computations[2].
- The reflex
- Know who owns the agent, who monitors its behavior and who can stop it[4]. This entry also advises alerting the responsible team if an agent seems to be running endlessly.
- What the company should plan for
- Microsoft recommends quotas, spending caps and cost alerts[2].
What it is
An AI service billed by usage costs money with every operation. Denial of Wallet (DoW) exploits that model: the attacker triggers a large number of operations, which creates unsustainable financial charges for the service provider, up to the risk of financial ruin[1]. OWASP files this case under LLM10:2025 Unbounded Consumption[1].
The service does not have to go down for the attack to succeed. Microsoft notes that an attacker can do harm without crashing the system: excessive use is enough to consume the budget, which is a financial denial of service[2]. This entry describes the mechanism, a case demonstrated in a laboratory in September 2026, and what an employee can do.
Straight answers
What is denial of wallet?
OWASP describes it as launching a large number of operations to exploit the pay-per-use model of cloud AI services[1]. OWASP's good-practice sheet on agents also describes it as attacks that cause excessive API or compute costs through unbounded agent loops[5].
Does the service go down?
Not necessarily. Microsoft explains that the attacker does not need to crash a system completely to cause damage[2]. When the service is saturated, however, legitimate users may experience slowness or unavailability[2].
Are many attacks of this kind seen?
Microsoft states that complete incidents of malicious denial of service against an LLM are not widely made public, but that clear demonstrations of the risk exist[2]. The dated case below is a laboratory study: the abstract describes an evaluation on a test bench and mentions no observed attack[3].
What does the company do to protect itself?
Microsoft recommends rate limits and quotas per user, IP address, API key, agent identity or service account, also applied to the tool calls an agent launches[2]. It also recommends tracking usage with dashboards and alerts, and setting spending limits on cloud AI services to avoid billing surprises[2]. OWASP adds limited input sizes, timeouts and monitoring of resource consumption[1].
What can an employee do?
The NCSC (UK) recommends that the organization know who owns an agent system, who approves its access, who monitors its behavior and who can stop it, and that these responsibilities be defined before the agent is connected to real systems[4]. For an employee, this entry draws one reflex from it: know which team is responsible for the agent they use, and alert that team when the agent behaves abnormally.
Inside the attacker's head
The attacker wants to make the service work as hard as possible, at the victim's expense. Microsoft cites concrete methods: sending a flood of requests to an internal assistant bot, or crafting prompts that produce very long or complex responses[2]. One example Microsoft gives is a request to produce the entire Bible in JSON format, repeated automatically[2].
Illustration (paraphrased prompt, no real target):
[Request to produce an enormous text, repeated by a script,
thousands of times, against an AI assistant billed per token]Microsoft also describes the case where the attacker offloads their own expensive computations onto the victim's model, which creates a hidden expense alongside classic denial of service[2]. It finally notes that it is hard to tell malicious use from normal but complex use, especially if the attacker mixes ordinary and heavy requests to slip under monitoring[2].
With an agent, the attacker has another lever: a single prompt can lead the agent to call APIs, create tasks or communicate with other agents at machine speed[2].
What you see on screen
The employee mostly notices the effects on the service: slow responses or unavailability[2]. Microsoft reports that in 2023, heavy demand for ChatGPT, with users submitting very long conversations to it, caused notable slowdowns and outages[2]. This case is not an attack, but it shows that heavy consumption degrades availability[2].
With an agent, the loop is noticed less quickly: the NCSC notes that problems are harder to spot when actions move faster than a human can review them[4]. Microsoft cites an early version of an autonomous GPT-4 agent that consumed exorbitant API credits by getting stuck in a loop of tasks it generated itself[2]. The spending shows up mostly in the dashboards and cost alerts that the company sets up[2].
Under the hoodIT teams
According to Microsoft, three mechanisms drive the load up. Very long inputs force the model to spend more time and memory. Pathological prompts ask, for example, to list a million prime numbers. Parallel requests, launched by scripts or bots, exceed the model's capacity[2]. OWASP also cites the continuous sending of inputs that exceed the model's context window[1].
Agents add a mechanism described by a study published on arXiv. An agent that calls tools keeps the state of the conversation from one turn to the next, and when a tool result is carried into later inputs, the provider bills it again[3]. The authors call this retained content persistent billable state[3]. A malicious or compromised tool, once admitted, can thus turn untrusted data into recurring processing billed to the victim, without having the victim's credentials or any local privilege[3].
On the defense side, OWASP's good-practice sheet asks to track token usage and costs per session and per user, and to impose limits on tokens, cost, retries and tool chaining[5]. It also asks for tests in which the limits on chain depth, retries, tokens and cost stop uncontrolled loops[5]. Microsoft recommends limiting the number of simultaneous agents, bounding recursion depth and capping tool calls[2].
Case filesReal cases
September 2026: a malicious tool inflates an agent's bill (laboratory study)
On 23 September 2026, researchers from the Institute of Information Engineering of the Chinese Academy of Sciences and other institutions filed a study on this risk on arXiv[3]. They describe six denial-of-wallet attack vectors and a test bench, DOW-BENCH, evaluated on six model families[3].
Across 243 runs, a session's cumulative input consumption reached at most 14,293 times that of the first call[3]. Keeping the raw history raises a session's average effective cost by 21.2 to 35.9%[3]. The authors propose four host-side safeguards that bound the request mass, context growth, recursion possibilities and cumulative spend before re-injection, and that contain every recurring attack in the corpus of 123 replayed evaluations[3]. Only 71 of the 3,830 MCP and transport server repositories analyzed expose a visible protection in the code, and none covers all four families[3].
This is a flaw demonstrated in a laboratory. The abstract describes an evaluation on a test bench and mentions no observed attack[3].
Sources
- OWASP, “LLM10:2025 Unbounded Consumption”, 10 April 2024. https://genai.owasp.org/llmrisk/llm102025-unbounded-consumption/ (accessed 14 November 2026)
- Microsoft, “5. Unbounded AI Consumption and Agentic DoS (DoS / Wallet Attack)”. https://learn.microsoft.com/en-us/security/zero-trust/catalog-ai-attack-techniques/unbounded-ai-consumption-and-agentic-dos (accessed 14 November 2026)
- arXiv, “Persistent Billable State: Denial-of-Wallet Attacks and Defenses in Tool-Calling LLM Agents”, 23 September 2026. https://arxiv.org/abs/2609.28585 (accessed 14 November 2026)
- NCSC (UK), “Thinking carefully before adopting agentic AI”, 15 May 2026. https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai (accessed 14 November 2026)
- OWASP, “AI Agent Security Cheat Sheet”. https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html (accessed 14 November 2026)
Cite this article
Brain, “Denial of wallet: draining an AI service”, Cyberpedia, 2026. https:/
BibTeX
@misc{brain_cp288_en_2026,
author = {Brain},
title = {Denial of wallet: draining an AI service},
year = {2026},
howpublished = {Cyberpedia},
url = {https://startbrain.io/cyberpedia/en/articles/denial-of-wallet-ai-service/},
note = {CC BY 4.0}
}