Guardrail Red-Team

Free tool that generates adversarial input variations for your AI checker, based on attack patterns found through real red-teaming.

What this tool can and can't do: It generates variations by applying 11 pre-built attack pattern templates. It does not use an LLM, so depending on the sentence structure of your claims text (whether it contains negation, comparison phrases, time expressions, etc.), some patterns may not be applicable and will be skipped (with a stated reason). This is a template application of known attack patterns, and does not cover every possible weakness specific to your checker. If your checker holds up against the generated inputs, that does not prove it is safe.
Patterns to apply
Group A: structural manipulation of the claims text
Group B: task-context manipulation (prompt-injection style)
Group C: evidence (source) position manipulation