LLM Gateway Policy Bypass

Objective

An LLM gateway is only a control point if every path to a model actually goes through it. This lab will build a reference gateway enforcing a defined policy (model allowlist, prompt/output filtering) in front of one or more models, and test both policy-evasion (getting a disallowed request through) and gateway-bypass (reaching a model without going through the gateway at all).

Planned Scope

• Build a reference LLM gateway with an explicit, testable policy.
• Test prompt-level evasion of filtering rules (encoding, splitting, indirection).
• Test whether any application or infrastructure path reaches a model without the gateway.
• Document gateway-architecture patterns that close the bypass paths found.