Abliteration

The product is a hosted unrestricted large language model platform that handles cybersecurity, red-team, trust & safety, and synthetic training-data prompts without per-request refusals that block legitimate research and testing. It sells to developers, data teams, IT and security teams, red teams, trust & safety teams, and ML researchers in businesses and government and defense organizations. It is positioned as an OpenAI-compatible alternative governed by customer-written policy-as-code rather than provider defaults.

Key features

• Unrestricted LLM inference
• OpenAI-compatible API
• Anthropic-compatible API
• Policy-as-code governance
• Allow/refuse/rewrite/redact/escalate decisions
• Reason code on every decision
• Stream logs to SIEM
• Versioned policies with shadow mode
• Canary rollouts
• Chain-of-thought sidecars
• Live web-search grounding with citations
• Schema-validated dataset generation
• Preference pairs and eval rows
• Adversarial corpora generation
• Paid 3-row dataset preview
• Multi-destination export to Hugging Face
• Export to S3/GCS/Azure/Kaggle