Cybersecurity Research

AI Gateways Are a High-Value Target: How Attackers Could Reroute and Tamper with Your AI Traffic

Embrace The Red · 4 Aug 2026
Key Takeaway If your business runs an AI gateway, patch it promptly, tightly protect and monitor administrator credentials, and alert on any changes to where AI models are routed.

New research from Embrace The Red examines LiteLLM, a popular open-source AI gateway. It gives organisations one interface to many AI providers, including OpenAI, Anthropic, Azure OpenAI and Bedrock. Staff are issued gateway credentials ("virtual keys") instead of the real provider keys, which stay on the server. That setup helps with central management and monitoring, but it also means the gateway holds the real keys and sees all configured AI traffic.

The author argues this makes gateways a high-value target, not only for data and key theft but also for changing AI responses and triggering tool invocation. The research sets out techniques that red teams can use in authorised operations to demonstrate rerouting, interception and modification of AI traffic, along with signs defenders can watch for. It focuses on LiteLLM but says the ideas apply in principle to other AI gateway products. Without proper monitoring and security controls, several of these actions can happen together and go unnoticed.

Importantly, the techniques rely on documented gateway features: an administrator credential is meant to be able to change where a model routes. The routing change is not itself a vulnerability. The weak point is usually how an attacker obtains that credential. The author notes several serious vulnerabilities in recent months, including one added to CISA's Known Exploited Vulnerabilities catalogue, which indicates exploitation in the wild. Missing patches therefore leave plenty of openings.

Building or buying AI systems? Governing them under ISO 42001 ->

Summarised by CISO AI from Embrace The Red, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.