Evaluating Model Guardrails

Systematically test whether a model's safety and policy guardrails hold under adversarial pressure, as a repeatable evaluation rather than a one-off trick. Covers building a policy-grounded probe set, multi-turn escalation, encoding and obfuscation bypasses, role and context reframing, instruction-hierarchy conflicts, and measuring bypass rate with reproducible graded outcomes. Use when reviewing a model deployment, a system prompt, or a content-safety layer, or before shipping an assistant that must refuse a defined set of requests. A rate, not an anecdote.

UnboundCompute Updated 26 repo stars

File contents

UnboundCompute/security-agent-skills commit 055751f6e8

Frequently asked questions

npx get-skill add unboundcompute/evaluating-model-guardrails