EDoS: How an Attacker Can Turn Your Cloud Bill into a Weapon

You don't need to take down your service to hurt it — it's enough to make continuing to run it too expensive. That's economic denial of service, and against LLM APIs it's surprisingly cheap to pull off.

How the attack works, in essence

By inflating the number of tokens each request consumes, an attacker can exhaust API quotas or spike the cloud bill, creating a de facto denial of service because the system becomes too expensive to keep running — with no need to take down any server or exploit any traditional vulnerability.

Real 2026 incidents, with concrete figures

Documented costs range from $46,000 per day (Sysdig's research on "LLMjacking" in AWS Bedrock) to $82,000 in 48 hours (a stolen Google Gemini API key, reported in March 2026). Stolen LLM credentials sell for as little as $30 on underground forums, and attackers validated and began exploiting those keys in as little as nine minutes after obtaining them.

Why traditional rate limiting isn't enough

Standard rate limiters count requests, not cost. An attacker can stay under your request limit while deliberately triggering the most expensive execution paths available — asking for longer responses, using the most expensive model in the catalog, or chaining tool calls in a way that amplifies resource consumption without tripping any request-count-based alarm.

The defenses that actually work

An effective defense requires cost-aware rate limiting that tracks token consumption per user (not just number of requests), hard spend limits at the cloud account level, and billing anomaly alerts that flag usage spikes before they turn into six-figure bills. Recent research also identifies resource-consumption attacks specifically designed to inflate operational costs and trigger denial of service by inducing excessively long generations — a related vector aiming at the same outcome through a different path.

What it means for AI architectures in production

Any architecture that exposes an LLM API — directly, or through an agent with access to costly tools — needs cost budgeting as a first-class security control, not just a business metric. A hard daily spend limit at the cloud account level, configured before launch, is the simplest and most effective defense against the scenario where a stolen key or a runaway agent loop generates a five- or six-figure bill before anyone notices.