Definition
An attack where a low-privilege user (or a low-privilege AI agent) exploits a vulnerability to gain elevated permissions, often leading to system admin access or the ability to inject malicious responses into an AI system. In AI gateways, privilege escalation can allow an attacker to rewrite model responses or exfiltrate all provider API keys.
Why it matters
AI gateways and agents often run with standing privileges under shared service accounts. A privilege escalation in an AI gateway doesn't just compromise that tool—it exposes every connected LLM provider, every API key, and every downstream system the agent can reach.