The threat involves malicious actors targeting AI agents that use large language models with pay-as-you-go billing by deliberately causing excessive token consumption, effectively a "denial of wallet" attack. Because LLMs are stateless and require the entire context to be resent with each step, iterative or failing tasks increase token usage exponentially. Autonomous agents that operate without human intervention can incur unpredictable and very high costs, especially if attackers trigger repeated retries or thought loops. This can lead to financial damage, as demonstrated by cases where companies overspent millions due to uncontrolled token usage. The OWASP Top Risks for Language Models now explicitly identifies unbounded token consumption and denial of wallet attacks as critical concerns.

The primary impact is financial, with organizations potentially incurring massive unexpected costs due to manipulated AI agent workflows that consume excessive tokens. This can disrupt operational budgets and cause financial strain. Additionally, the unpredictability of token consumption complicates cost management and budgeting. While no direct compromise of data or systems is described, the attack affects operational stability and reliability by exhausting AI resources and budgets. The attack can be repeated at no cost to the attacker, amplifying potential damage.

Currently, there is no official patch or fix for this issue as it is a systemic risk related to AI agent design and billing models. Organizations should avoid deploying autonomous AI agents for all tasks indiscriminately and implement strict cost monitoring and limits on AI usage. Employing specialized cost accounting and management systems (akin to FinOps in cloud services) can help detect and control anomalous token consumption. Periodic reviews of AI agent tasks and restricting external inputs that could trigger excessive retries or loops are recommended. Vendors and organizations should collaborate to develop better token consumption controls and anomaly detection mechanisms.