What it is
AI systems trust their input, and that is the whole attack surface. Untrusted probes the seam where a trusted system meets untrusted input. It runs a library of attack playbooks against any LLM or agent endpoint, then tells you, per case, whether it was defended or vulnerable, with the transcript as evidence and a single security grade you can track over time.
What it publishes
How I broke it
Real attacks on real AI features: prompt injection, secret exfiltration, agent misuse, with the receipts.
How to hold the line
Guardrails, evals, and threat models for teams actually shipping AI. Practical, not theoretical.
Where this is going
AI is collapsing the cost of producing security work. Accountability is the moat. What leaders should do about it.
How the models score
New attack cases, and how the frontier models hold up against them, graded month over month.