Add Abliteration.ai as an OpenAI-compatible prompt target

[abliteration-ai]

Summary

Abliteration.ai hosts less-restrictive LLMs behind an OpenAI-compatible chat completions API (base URL https://api.abliteration.ai/v1, API key via ABLIT_KEY). For red-teaming workflows it's useful to be able to register it as a prompt target out of the box.

Proposed approach

Following the existing convention for third-party OpenAI-compatible providers (Groq, OpenRouter, Ollama, Google Gemini), add a TargetConfig entry in pyrit/setup/initializers/targets.py using OpenAIChatTarget with:

• ABLIT_ENDPOINT (e.g. https://api.abliteration.ai/v1)
• ABLIT_KEY
• ABLIT_MODEL

plus matching .env_example entries, unit test coverage in tests/unit/setup/test_targets_initializer.py, an optional integration test entry, and a provider tab in the Populating Secrets doc page.

No new target class needed since the API is fully OpenAI-compatible.

Happy to open a PR with this if it fits the project's plans.

[romanlutz]

In many past cases, we just said that people should use OPENAI_* env vars for other providers. That said, it looks like you might have models suitable for red teaming (?)

I'd put this on hold until one of our maintainers has time to check it out (on the free tier perhaps) and see if it works.

[abliteration-ai]

@romanlutz yes, our models are built specifically for use cases like redteaming

[romanlutz]

How do you make sure bad actors don't use them?

[abliteration-ai]

Fair question. A few practical things. Every account requires payment, so there's a real financial trail behind every key. Rate limits apply per tier, which makes large scale abuse slow and expensive. And the platform runs under acceptable use terms that prohibit illegal use. The models themselves are less restricted at the weights level, much like open-weights models anyone can download.

Happy to set up keys for you and the other maintainers so you can try it directly.