security-testing

Building an AI that pentests your AI, with proof required

An engineer built an adversarial LLM harness that only reports proven exploits against a live agent, replacing vague 'might be vulnerable' pentest reports with deterministic, oracle-verified findings.