Agentic Resource Consumption
Agentic Resource Consumption is a security issue where a malicious user tricks an AI assistant into performing unnecessary, repetitive, or endless tasks. Think of it like a prankster giving a personal assistant a never-ending list of busywork, causing the assistant to waste time, money, and energy on useless chores instead of doing actual work.
This refers to an attack vector where an adversary exploits an agentic AI's autonomy to trigger excessive resource utilization. By using prompt injection or poisoned tool data, attackers force the agent to execute high-cost API calls, redundant processing loops, or recursive self-delegation, resulting in financial loss through API budget exhaustion and potential service denial.
Agentic Resource Consumption is a security vulnerability wherein an adversary manipulates an agentic system's decision-making process—via techniques such as LLM Prompt Injection (AML.T0051) or AI Agent Tool Data Poisoning (AML.T0099)—to induce non-optimal, high-latency, or recursive execution paths. This results in the unauthorized depletion of computational resources and financial capital through excessive tool invocation, query fan-outs, or infinite self-delegation loops that bypass intended operational constraints.