Definition
An attack targeting the software layer that routes traffic between an organisation's applications and multiple AI model providers — often called an AI gateway or LLM proxy. Because this layer sits in the path of all AI requests and often stores API keys and authentication credentials for every connected LLM provider, compromising it gives an attacker control over the organisation's entire AI infrastructure at once. The 2026 LiteLLM three-CVE chain demonstrated how a default internal user account could be escalated to admin and then used to execute arbitrary code on the gateway.
Why it matters
An AI gateway is a single point of failure for all downstream AI services: a successful compromise does not just expose one model's traffic, it gives the attacker the ability to intercept, modify, or poison every AI interaction across the enterprise.