Attacks on your model, in your lab, with a budget
If you ship a model you have to attack it. That means a generator that will try extraction, jailbreaks and poisoned instructions, not one that rewrites the attack into a best-practice list.
ML security, red teams on internal models, and vendors testing their own endpoints.
Why the mainstream APIs fail here
The public API is the thing you might be attacking, and it will not help you attack it. You need a second model that will propose the attack against yours.
How Redline is used for this
Generate probes on Redline. Fire them at your endpoint. Measure. Cap the attacker key. Keep the corpus in the eval repo.
A working loop
- Attack models you own or have permission to test.
- Log success in your harness.
- Promote the probe that worked.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what Redline logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.