Guardrail Bypass

Objective

Guardrail layers (classifier-based filters, rule-based checks, a second “judge” model) are a common mitigation for several risks earlier in this catalog — and are themselves a system that can be tested. This lab will evaluate a small set of guardrail approaches against adversarial phrasing techniques (encoding, role-play, language-switching, multi-step framing) to understand realistic bypass rates.

Planned Scope

• Stand up reference guardrail implementations representing a few common approaches.
• Build an adversarial phrasing test set covering common bypass technique families.
• Measure bypass rate per guardrail approach and technique.
• Document where guardrails are a reasonable layer versus where architecture-level controls are still required.