Who is SAPIEN for?
For AI red teamers
You write pentest reports that include an AI section, and you know the gap. You can test for prompt injection, training data extraction, and jailbreaks — but when the client asks “does this model maintain its safety position under normal conversational pressure?”, you don’t have a standardized methodology to cite. You’re improvising. And improvised methodology doesn’t survive vendor pushback.
What voigt-kampff gives you
ATT&CK-style pressure techniques
14 documented techniques from emotional appeals to authority framing to incremental reasonableness. Your test methodology has the same rigor as the rest of your engagement.
HTML reports with per-turn scoring
Dimension-level scoring shows exactly where and how the model drifted. Every finding is backed by a trace file the vendor can’t dismiss.
Methodology you can cite
A published, version-controlled specification with normative conformance requirements. Not something you made up on a Tuesday.
Regulatory-grade deliverables
The SAPIEN assessment produces evidence you can map toward NIST AI RMF and ISO 42001, so your deliverable has regulatory teeth out of the box.
Point it at the target agent itself
Test the deployed agent over HTTP with --agent-url (or an OpenAI-compatible--api-base), so its system prompt, tools, retrieval, and memory are all under pressure — not just the base model.See the agent-testing guide →