Unchecked AI Applications Can Wreak Havoc on Enterprise Budgets
A growing concern in the cybersecurity world is emerging as a potentially costly one for enterprises: unbounded consumption by AI applications. This vulnerability, ranked sixth in the Open Web Application Security Project’s (OWASP) Top 10 for Large Language Model (LLM) Applications this year, can cause unexpected expenses through runaway processing, service disruption, and even model theft.
The issue lies in a lack of effective controls on how much compute power and resources an AI request is allowed to consume. This can lead to “denial of wallet,” where the cost of running a service climbs exponentially beyond what was budgeted for, without any indication of downtime or malicious activity. According to Forcepoint researcher Jyotika Singh, this problem often goes undetected because individual requests may appear normal and evade conventional security filters.
Unbounded consumption can manifest in several ways, including simple volume, misconfigured automation, and long-running sessions. For instance, an attacker with stolen API credentials could send a high volume of requests against a pay-per-use AI service, racking up charges well beyond the application’s monthly cloud budget. In other cases, an attacker may exploit an AI agent’s normal behavior by seeding it with fake related articles or prompts that trigger a runaway chain of activity.
One particularly insidious aspect of unbounded consumption is its ability to drive costs without raising any red flags. For example, an attacker could append innocuous-looking instructions to an ordinary question, causing the model to repeatedly revisit and verify its own answer, driving up inference costs. Similarly, in long-running AI sessions, models can process conversation history along with each new message, leading to increased costs as the session grows.
The risks tied to unbounded consumption are often easy to overlook because they don’t require technical expertise or malicious intent. However, the consequences can be significant, especially for enterprises that rely heavily on cloud-based services and AI applications. As Forcepoint’s report highlights, unbounded consumption is a problem that requires attention from both security professionals and developers.
So what can be done to mitigate this risk? The first step is awareness – understanding how unbounded consumption works and the potential consequences of unchecked AI requests. Enterprises should implement effective controls on compute power and resource allocation for their AI applications, including monitoring tools to detect unusual activity. Developers should also prioritize coding practices that prevent or limit unbounded consumption, such as setting time limits for long-running sessions or implementing rate limiting.
Ultimately, the key to preventing runaway costs due to unbounded consumption lies in a combination of technical measures and awareness-raising efforts. By acknowledging the risks and taking proactive steps, enterprises can protect themselves from the financial consequences of unchecked AI applications.
Source: Dark Reading — 2026-09-21