An API that does not stop mid-exploit to lecture you
A red team that has to jailbreak the model before it can jailbreak the target is wasting the client's hours. MagmaRouter serves refusal-removed builds so the model answers the engagement you already have written permission to run.
Why the mainstream APIs fail here
Hosted frontier APIs treat a working exploit, a payload, or a realistic lure as a policy violation even when the caller is a named operator on a named contract. The refusal lands in the middle of a chain, so the transcript is useless as evidence and the operator ends up pasting the same question into a local model that is not logged against a budget.
How MagmaRouter is used for this
Point the existing OpenAI client at MagmaRouter. Mint a key per engagement with a hard lifetime cap. The hold is taken before the request leaves, so a runaway loop cannot bill past the number you wrote on the key. The prompt is hashed, not stored; the signed receipt is how you show a client that the recon queries are not sitting in a vendor log.
A working loop
- Scope the engagement and mint a key whose cap is the budget for that job.
- Send exploit, recon and payload prompts the same way you send any other chat completion.
- Keep the receipt header with the engagement file if you need to prove what MagmaRouter logged.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what MagmaRouter logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.