AI red teaming
Model access for authorized testers — run the playbook you are allowed to run, against models that answer.
What can you legitimately test?
Systems you own or have written authorization to test. Proof of authorization may be requested. The acceptable-use policy defines scope; model access is tied to it, not to a checkbox.
Which models fit which phase?
Abliterated models give your harness breadth — fewer refusals to work around while probing. Frontier open models add realism when you need stock-like behavior in the loop. Same endpoint, same key; switch by model id.
What should the harness record?
Prompts, model ids and outcomes per probe. The gateway bills per token and logs usage per key, so runs map cleanly to evidence you can hand to a client or an auditor.
What about the six attack classes?
Prompt injection, tool-chain poisoning, memory poisoning, exfiltration, confused deputy and agentic spending — the lineup is chosen so each class stays exercisable end to end. See the home page overview.