AI systems can generate unexpectedly high costs when organisations do not limit the compute, tokens and other resources available to each request. Dark Reading reports that Forcepoint researcher Jyotika Singh described this as “unbounded consumption”, ranked sixth in OWASP’s 2026 Top 10 for LLM Applications. The service may remain available while costs rise beyond budget, a pattern known as “denial of wallet”.
The risk can result from stolen API credentials, misconfigured automation, long-running sessions or ordinary-looking requests, and does not always require a technically skilled or malicious actor.
Forcepoint identified five scenarios: attackers using leaked keys to send large volumes of paid requests; AI-agent tool fan-out, where poisoned content causes an agent to follow expanding chains of links; reasoning-loop exhaustion, which makes models spend excessive “thinking tokens”; context accumulation, where long sessions repeatedly process growing histories; and model extraction through large numbers of queries to an exposed inference endpoint. Singh gave the example of a support session lasting more than 150 exchanges, with per-message costs rising roughly 100-fold by turn 100.
The recommended response is to impose hard spending and token limits for users, API keys and teams, rather than relying solely on alerts. Organisations should also cap agent steps and self-repetition, detect loops, and use sandboxing and least-privilege controls to restrict access and contain runaway activity.