AI Token Consumption Spirals Out of Control: The Rise of "Denial of Wallet" Attacks In 2026, enterprises faced an unexpected crisis as AI adoption outpaced cost controls, leading to runaway spending and a new cybersecurity threat: denial-of-wallet attacks. Companies like Uber exhausted their annual AI budgets by April, while one unnamed organization racked up a $500 million bill in a single month after failing to set spending limits on Anthropic’s Claude. The shift from fixed-subscription models to pay-as-you-go billing coupled with the rise of autonomous AI agents has turned token consumption into a financial and security liability. ### Why Costs Are Spiraling Unlike traditional machine learning (ML) or human-operated chatbots, autonomous AI agents operate continuously, resending entire task histories (context) with each step. Errors, retries, and iterative loops cause token usage to balloon unpredictably sometimes by 30x or more for identical tasks. A single agent stuck in a "thought loop" can waste vast resources on trivial work, while multi-agent systems compound the problem by multiplying context exchanges. ### The New DDoS: Denial of Wallet Attackers are exploiting this unpredictability to financially cripple organizations. By flooding AI systems with malicious or overly complex requests, they trigger excessive token consumption. For example: - GitInject attacks on GitHub could cost victims $111 per incident and burn 400 minutes of GitHub Actions before defenses activate. - "OverThink" exploits demonstrated how a benign prompt could inflate token usage by 46x while bypassing security filters. - Gartner estimates a single LLM-powered support request costs $3 making mass-generated attacks a low-effort, high-impact threat. The OWASP Top 10 for LLMs (2026) now ranks unbounded token consumption (LLM06) as a top risk, explicitly warning of denial-of-wallet attacks that drain budgets before anomalies are detected. ### The Root of the Problem Three generations of AI systems consume resources differently: 1. Classical ML: Predictable, low-cost, fixed budget. 2. LLM chatbots: Costs scale with user activity, manageable via licenses. 3. Autonomous agents: No cost ceiling token usage grows exponentially with task complexity, errors, and retries. Without FinOps-style cost controls (common in cloud and telecom), companies only discover the true cost of AI processes after the fact. The probabilistic nature of generative AI further complicates forecasting, leaving organizations vulnerable to both accidental overspending and targeted attacks. The shift to AI agents has exposed a critical gap: enterprises lack the tools to monitor, predict, or cap token consumption making them prime targets for a new breed of financial sabotage.