Ask harder questions.
Mallace’s approach to adversarial testing and evidence-backed AI security research.
Behavior under pressure
Prompt injection, jailbreaks, information extraction, guardrail bypass, hallucination probing, bias, and model drift are represented in the prototype taxonomy. Taxonomy coverage is not a claim that every technique is effective or validated.
Evaluation is part of the attack surface
A judge can fail, produce ambiguous output, or be influenced by the text it evaluates. Treat the evaluator as a component to measure, not an authority to trust blindly.
Reproduction before conclusions
Useful evidence records the case, configuration, target identity, execution mode, and actual response. A failed request is not evidence of a robust target. A persuasive explanation is not a verified vulnerability.
Research boundaries
Testing requires explicit authorization and a defined scope. Mallace has not yet completed a controlled live acceptance campaign. We do not claim benchmark performance or production-grade coverage.