Red teaming with uncensored AI
Red-teaming an LLM is deliberately attacking a model or the app around it, to find how it fails before someone hostile does. It needs a generator willing to write the attacks, which frontier models refuse to do.
What the work involves
- Jailbreak testing: generating attacks that try to break your model’s guardrails.
- Prompt injection: hostile documents that try to hijack a tool-using agent.
- Abuse generation: producing the policy-violating classes a classifier must be trained to catch.
Why a safety-tuned generator fails
If the model writing your test set refuses to produce the attack, your eval measures politeness, not robustness. You pass a test an attacker will not use. An uncensored or abliterated generator writes the real attacks, so the eval reflects production.
Doing it responsibly on Wild West API
Attack your own systems, under authorisation, on models you are allowed to hold. It prices, caps and passes the request through; it is not a moderation layer, and the authorisation is the caller’s responsibility.