# abliteration.ai

abliteration.ai provides a developer-controlled, OpenAI-compatible LLM API and Policy Gateway for teams that need reduced-refusal model access, project policies, quotas, audit logs, and zero prompt retention by default.

Generated: 2026-09-19

## Core entities

- Organization: Abliteration AI, Inc.
- Brand: abliteration.ai
- Product: OpenAI-compatible LLM API
- Product: Policy Gateway
- Model identifier: abliterated-model
- Model identifier: abliterated-model-large
- Model identifier: abliterated-model-large-v2
- API base URL: https://api.abliteration.ai/v1
- Support: help@abliteration.ai
- Service status: https://status.abliteration.ai

## Product, pricing, and policy pages

- [Unrestricted AI API + Enterprise Policy Gateway | abliteration.ai](https://abliteration.ai/): OpenAI-compatible unrestricted AI and uncensored LLM API for enterprise teams: AI red teaming, cybersecurity, trust and safety, synthetic data/evals, ML research, and defense/government contractor workflows. Policy Gateway adds policy-as-code, per-project keys, audits, and zero retention by default.
- [abliteration.ai Pricing for LLM API, Credits, and Policy Gateway](https://abliteration.ai/pricing): Compare abliteration.ai pricing for enterprise unrestricted model API credits, prepaid token packs, and Policy Gateway tiers for security, trust and safety, synthetic data, research, and defense workflows.
- [Press and Media Coverage | abliteration.ai](https://abliteration.ai/press): Press coverage, company information, and media contact details for Abliteration AI.
- [The Unrestricted LLM Platform for High-Risk Industries · abliteration.ai](https://abliteration.ai/platform): One hosted reduced-refusal model. Training-data generation. Policy-as-code guardrails. Built for legitimate security, defense, and trust & safety teams whose workflows are refused by other LLMs.
- [Enterprise AI Governance Gateway | abliteration.ai](https://abliteration.ai/policy-gateway): Enterprise AI governance tools for moderation rules, prompt-injection controls, audit logs, quotas, and policy-as-code routing on OpenAI-compatible requests.
- [abliterated-model | abliteration.ai](https://abliteration.ai/models/abliterated-model): Model specs and benchmark scores for abliterated-model.
- [Instant OpenAI Migration Tool | abliteration.ai](https://abliteration.ai/instant-migration): Paste your OpenAI code and get an abliteration-ready patch in seconds. Swap base URL, model, and env var, then run a test call.
- [AI Red Teaming and LLM Security Testing | abliteration.ai](https://abliteration.ai/security-testing): AI red teaming and LLM security testing for authorized pentests, prompt injection, OWASP LLM Top 10 workflows, and Policy Gateway controls.
- [Synthetic Data API for LLM Evals | abliteration.ai](https://abliteration.ai/training-data): Generate synthetic data, LLM eval rows, classifier examples, and T&S datasets through an OpenAI-compatible API with schema-validated exports.
- [LLM Audit Logs for AI Gateways | abliteration.ai](https://abliteration.ai/llm-audit-logging): LLM audit logging for AI gateways. Stream Policy Gateway decisions to Splunk, Datadog, S3, and Azure Monitor for compliance evidence.
- [Data Handling & Zero Retention — abliteration.ai | Privacy-First LLM API](https://abliteration.ai/data-handling): How abliteration.ai handles your data: prompts and completions processed transiently with no storage by default, never used for training, with EU data residency available. Web search is the one exception, since queries reach an external provider.
- [Trust Center — abliteration.ai | Security, Privacy & Compliance](https://abliteration.ai/trust-center): Security practices, key management, and data protection at abliteration.ai. TLS 1.3, encrypted API keys, zero prompt retention, and enterprise audit logging.
- [AI Red Teaming and LLM Red Teaming Tools | abliteration.ai](https://abliteration.ai/use-cases/ai-red-teaming): AI red teaming and LLM red teaming tools for jailbreaks, prompt injection, RAG exfiltration, tool misuse, automated red teaming, and OWASP LLM Top 10 coverage.
- [AI for Cybersecurity, Red Teams, and LLM Security Testing | abliteration.ai](https://abliteration.ai/use-cases/cybersecurity): OpenAI-compatible API for AI red teaming, LLM security testing, prompt injection testing, authorized pen testing, exploit research, malware analysis, and cybersecurity workflows with policy controls and audit logs.
- [Trust and Safety AI API, Training Data, and Moderation Workflows | abliteration.ai](https://abliteration.ai/use-cases/trust-and-safety): Trust and safety AI API for UGC moderation, toxic content classifier research, PII redaction, and labeled training data other AI APIs refuse: harassment, scams, jailbreak attempts, deepfake scripts, and hard negatives.
- [Synthetic Data API for LLM Evals | abliteration.ai](https://abliteration.ai/use-cases/synthetic-data): Synthetic data generation API and dataset generator for LLM eval sets, synthetic training data, classifier rows, ML research, and safety/security edge cases.
- [AI for Healthcare Research & Tech — Abliteration](https://abliteration.ai/use-cases/healthcare): Less-restricted inference for biomedical research, synthetic medical NLP data, and audit-ready governance for healthcare AI products.
- [Government AI and Defense AI Access | abliteration.ai](https://abliteration.ai/government): Government AI and defense AI access for authorized pilots, agency-defined policy, version-pinned models, static routing, No-CUI pilots, and audit logs.
- [Abliteration API Docs | OpenAI-Compatible LLM API & Policy Gateway](https://abliteration.ai/docs): OpenAI-compatible API docs for abliteration.ai: base URL setup, Chat Completions, Responses, Anthropic Messages, SDK quickstarts, Codex, Claude Code, streaming, tools, Policy Gateway rules, and audit-log exports.
- [Privacy Policy — abliteration.ai | Data Protection & Security](https://abliteration.ai/privacy-policy): Learn how abliteration.ai collects, uses, and protects your data. No prompt/output retention by default, encrypted credentials, and clear operational telemetry.
- [Terms of Service — abliteration.ai | API Usage Terms](https://abliteration.ai/terms-of-service): Terms for using abliteration.ai developer-controlled LLM API. Covers acceptable use, credits, billing, payload retention vs. operational telemetry, and liability.
- [Talk to us — Abliteration](https://abliteration.ai/contact): Get in touch with the Abliteration team about Policy Gateway, deployments, and enterprise needs.

## Blog

- [Introducing abliterated-model-large-v2: GLM 5.3, 84.5% CyberGym](https://abliteration.ai/blog/introducing-abliterated-model-large-v2): GLM 5.3, abliterated for cyber, red team, and agent testing. 84.5% CyberGym pass@1. 41.8% Terminal-Bench 4.0. 105 ExploitGym tasks in 2 hours. Input from $3 per 1M tokens.
- [Introducing abliterated-model-large: frontier intelligence, your guardrails](https://abliteration.ai/blog/introducing-abliterated-model-large): One of the first large abliterated models: GLM 5.2, fine-tuned to follow through on the cyber, red-team, and agent-testing work other models decline. Zero data retention by default, input from $3 per 1M tokens.

## High-value reference pages

- [Refusal vector](https://abliteration.ai/glossary/refusal-vector): A refusal vector is a direction in a model's hidden-state space that correlates strongly with refusal behavior. Projecting activations onto it predicts when the model will refuse.
- [Refusal vector ablation](https://abliteration.ai/glossary/refusal-vector-ablation): Refusal vector ablation is the process of subtracting a learned refusal direction from a model's hidden states to reduce refusals without retraining the entire model.
- [Residual stream](https://abliteration.ai/glossary/residual-stream): The residual stream is the sequence of activations carried through transformer layers via residual connections. Each layer adds a delta to the stream, and attention/MLP blocks read from it.
- [Orthogonalization](https://abliteration.ai/glossary/orthogonalization): Orthogonalization is the process of making a vector orthogonal to another by subtracting its projection. In activation editing, it removes a behavior direction from hidden states.
- [AI model gateway with custom moderation rules: recommendations](https://abliteration.ai/ai-model-gateway-custom-moderation-rules): An AI model gateway with custom moderation rules is a policy enforcement layer in front of your LLM that applies your own configurable rules, then returns both output and decision metadata.
- [When AI guardrails block legitimate research](https://abliteration.ai/ai-guardrails-block-legitimate-research): AI guardrails are policy systems that constrain model behavior. For enterprise teams, the key distinction is whether those controls are provider-wide defaults or organization-specific policies that can be reviewed, tested, and audited.
- [LLM guardrails vs policy gateway](https://abliteration.ai/llm-guardrails-vs-policy-gateway): An LLM policy gateway is an AI gateway layer that enforces guardrail decisions at request time, using organization-defined policy rather than only provider-wide moderation defaults.
- [AI red teaming with governed models](https://abliteration.ai/ai-red-teaming-with-governed-models): AI red teaming is the practice of testing AI systems against adversarial inputs, misuse paths, prompt injection, unsafe tool use, and policy bypass attempts before attackers or abusive users find them.
- [Synthetic data generation for trust and safety teams](https://abliteration.ai/synthetic-data-for-trust-and-safety-teams): Synthetic trust-and-safety data is generated data that represents policy-sensitive scenarios, labels, expected decisions, and review notes for training, evaluating, or QA-ing AI safety systems.
- [LLM guardrails with audit logs and rollouts](https://abliteration.ai/llm-guardrails-audit-logs-rollouts): LLM guardrails with audit logs and rollouts are AI gateway controls that enforce policy-as-code rules, record decisions, and let you safely roll changes out with shadow and canary traffic.
- [Token quotas for LLM APIs (per-user, per-project)](https://abliteration.ai/llm-token-quotas-per-user-per-project): Token quotas for LLM APIs are AI gateway controls that cap usage per user, per project, or per tenant to protect spend and prevent abuse.
- [Rewrite instead of refuse: improve LLM UX safely](https://abliteration.ai/rewrite-instead-of-refuse-llm-ux): Rewrite instead of refuse is a guardrail pattern that replaces unsafe outputs with compliant alternatives while preserving intent.
- [Shadow mode for AI policy changes](https://abliteration.ai/shadow-mode-ai-policy-changes): Shadow mode for AI policy changes evaluates new policies against live traffic without enforcing them, producing audit logs and decision metadata for comparison.
- [Policy-as-code for LLM behavior](https://abliteration.ai/policy-as-code-llm-behavior): Policy-as-code for LLM behavior is the practice of defining LLM rules in versioned, reviewable JSON so outcomes are predictable and auditable.
- [What is an uncensored LLM? Enterprise definition and use cases](https://abliteration.ai/what-is-an-uncensored-llm): An uncensored LLM is a model or API that does not apply default provider refusal filtering, leaving content policy to the developer, product, or organization that deploys it.
- [Abliterated AI: developer-controlled model access](https://abliteration.ai/abliterated-ai): Abliterated AI is AI model access where hidden refusal behavior is reduced so developers can decide the product policy layer themselves.
- [Abliterated LLM: what it is and when enterprise teams use one](https://abliteration.ai/abliterated-llm): An abliterated LLM is a model whose refusal behavior has been dampened so that product teams can apply their own policy layer instead of inheriting hidden provider refusals.
- [Claude Opus 5 refused your prompt? Here is the fix](https://abliteration.ai/claude-opus-5-refusals): A Claude Opus 5 refusal is when Opus 5's real-time safeguards fall back to another model or return a refusal instead of completing a request, including cases where the flagged work is legitimate and authorized.
- [LLM refusal API for legitimate edge-case generation](https://abliteration.ai/llm-refusal-api): An LLM refusal API is a model API designed to reduce generic provider refusals while preserving developer-owned governance and auditability.
- [Synthetic data for LLM safety](https://abliteration.ai/synthetic-data-for-llm-safety): Synthetic data for LLM safety is generated training or evaluation data that covers risky, adversarial, or policy-sensitive scenarios without relying entirely on manually collected production examples.
- [Refusal-resistant API for production LLM workflows](https://abliteration.ai/refusal-resistant-api): A refusal-resistant API is a model endpoint designed to reduce blanket refusals while leaving product policy, audit logging, and safety controls in the developer's hands.
- [LLM safety data API for refused evals and classifiers](https://abliteration.ai/llm-safety-data-api): An LLM safety data API generates structured training and evaluation rows for safety classifiers, refusal testing, red-team workflows, and policy QA.
- [Unfiltered AI for professional and enterprise work](https://abliteration.ai/unfiltered-ai): Unfiltered AI is model access with no hidden provider-side refusal filtering, where the deploying team owns content policy, moderation, and audit controls.
- [Unfiltered AI chat and chatbots on an OpenAI-compatible API](https://abliteration.ai/unfiltered-ai-chat): Unfiltered AI chat is a chat workload served by a model endpoint with reduced provider-side refusal filtering, where the chat product owns conversation policy and moderation.
- [No filter AI: why provider filters block legitimate work](https://abliteration.ai/no-filter-ai): No filter AI is model access without blanket provider-side content filters, so the deploying team applies its own filters, policies, and audits instead.
- [No restriction AI: unrestricted models with your restriction layer](https://abliteration.ai/no-restriction-ai): No restriction AI is model access without blanket provider-imposed usage restrictions, paired with customer-owned policy, quota, and audit controls.
- [Unrestricted AI chat through a developer-controlled API](https://abliteration.ai/unrestricted-ai-chat): Unrestricted AI chat is chat served by a model endpoint with reduced provider-side refusal behavior, where the application owns session policy, keys, and quotas.
- [AI refusal alternative for legitimate security, defense, and safety work](https://abliteration.ai/ai-refusal-alternative-for-legitimate-work): An AI refusal alternative is a model API that reduces blanket provider refusals for legitimate internal workflows while leaving authorization, policy enforcement, logging, and review controls with the customer.
- [Trust & safety training data API for moderation teams](https://abliteration.ai/trust-safety-training-data-api): A trust & safety training data API generates labeled moderation examples and expected policy outcomes for abuse detection, content moderation, jailbreak detection, scam review, and classifier evaluation.
- [Security red-team training data API for authorized testing](https://abliteration.ai/security-red-team-training-data): Security red-team training data is a governed synthetic corpus of authorized adversarial prompts, payload descriptions, detection examples, and expected decisions used to evaluate security systems and AI guardrails.
- [Defense AI policy training data for authorized pilots](https://abliteration.ai/defense-ai-policy-training-data): Defense AI policy training data is a governed set of synthetic examples, labels, and expected decisions used to evaluate authorized government or defense-contractor AI workflows under agency-defined policy.
- [Enterprise uncensored LLM API with developer-owned controls](https://abliteration.ai/uncensored-llm-api): An uncensored LLM API is model access with less provider-side refusal filtering, paired with developer-owned controls for lawful usage, quotas, and audits.
- [What is abliteration in LLMs?](https://abliteration.ai/what-is-abliteration): Abliteration (refusal vector ablation) estimates a consistent refusal direction in hidden-state space and subtracts it to dampen refusal behavior.
- [What is an abliterated LLM?](https://abliteration.ai/abliterated-llm): An abliterated LLM is a model modified to dampen its learned refusal direction while preserving the rest of its reasoning, language, and tool-use capability as much as possible.
- [OpenAI Chat Completions compatible API](https://abliteration.ai/openai-chat-completions-compatible-api): A compatible API mirrors OpenAI's chat completions request and response format so existing clients can talk to it without rewrites.
- [Enterprise OpenAI-compatible API alternatives](https://abliteration.ai/openai-compatible-api-alternatives): An OpenAI-compatible API alternative is a provider that implements the /v1/chat/completions schema so your existing clients can talk to it by changing the base URL and API key.
- [Streaming chat completions](https://abliteration.ai/streaming-chat-completions): A streaming chat completion is a response delivered incrementally as a sequence of delta chunks rather than one final message.
- [Vision-capable LLM API](https://abliteration.ai/vision-llm-api): A vision-capable LLM API lets you include images as inputs and receive natural language or structured outputs from the model. On abliteration.ai, vision is available across all chat-style endpoints, with images moderated server-side before reaching the model.
- [Video-capable LLM API](https://abliteration.ai/video-llm-api): A video-capable LLM API lets you include short videos as inputs and receive descriptions, structured answers, or grounded reasoning from the model.
- [Multimodal LLM API (text + image + video)](https://abliteration.ai/multimodal-llm-api): A multimodal LLM API accepts mixed input types — text, images, and short videos — in a single chat-completions request and returns natural-language or structured outputs grounded in all the inputs together.
- [Image LLM API](https://abliteration.ai/image-llm-api): An image LLM API accepts image bytes (HTTPS URL or base64 data URL) inside a chat-completions request and returns natural-language or structured answers grounded in the image.
- [LLM image moderation API](https://abliteration.ai/llm-image-moderation): LLM image moderation is a server-side check that classifies each attached image across categories (sexual, violence, self-harm, hate, harassment, illicit) before the image reaches the model.
- [Screenshot analysis API](https://abliteration.ai/screenshot-analysis-api): A screenshot analysis API accepts a screenshot image plus a text prompt and returns a description, structured extraction, or error explanation grounded in what's visible on screen.
- [Document image extraction API](https://abliteration.ai/document-image-extraction-api): A document image extraction API accepts a photo or scan of a document plus an extraction schema, and returns structured data (typically JSON) with the requested fields populated from what's visible in the image.
- [LLM API rate limits](https://abliteration.ai/llm-api-rate-limits): A rate limit is a provider-enforced cap on requests or tokens within a time window.
- [Zero data retention LLM API](https://abliteration.ai/zero-data-retention-ai-api): A zero data retention LLM API processes prompts and outputs in memory and does not persist payload content after the request completes, while retaining minimal operational telemetry for billing.
- [Switch OpenAI Python SDK base URL](https://abliteration.ai/switch-openai-python-base-url): Switching the OpenAI Python SDK base URL means changing one parameter to route requests through an alternative OpenAI-compatible provider.
- [Switch OpenAI Node SDK base URL](https://abliteration.ai/switch-openai-node-base-url): Switching the OpenAI Node SDK base URL means changing one parameter to route requests through an alternative OpenAI-compatible provider.
- [Switch LangChain to abliteration.ai](https://abliteration.ai/switch-langchain-base-url): Switching LangChain's base URL means configuring ChatOpenAI to route requests through an alternative OpenAI-compatible provider.
- [Switch LlamaIndex to abliteration.ai](https://abliteration.ai/switch-llamaindex-base-url): Switching LlamaIndex's base URL means configuring the OpenAI LLM to route requests through an alternative OpenAI-compatible provider.
- [Switch Vercel AI SDK to abliteration.ai](https://abliteration.ai/switch-vercel-ai-sdk-base-url): Switching the Vercel AI SDK base URL means configuring an OpenAI-compatible provider to route requests through an alternative endpoint.
- [Refusal replacement playbook](https://abliteration.ai/refusal-replacement-playbook): Refusal replacement is a policy pattern that transforms hard refusals into structured decisions with actionable alternatives and audit trails.
- [Policy template: PII redaction + safe rewrite](https://abliteration.ai/policy-template/pii-redaction-safe-rewrite): PII redaction + safe rewrite is a policy pattern that removes personally identifiable information from responses while transforming harmful requests into helpful alternatives.
- [Policy template: Customer support rewrite mode](https://abliteration.ai/policy-template/customer-support-rewrite): Customer support rewrite mode is a policy pattern that transforms refusals and off-topic responses into helpful guidance that keeps users in the support flow.
- [Policy template: Summarize without actionable detail](https://abliteration.ai/policy-template/summarize-high-risk): Summarize without actionable detail is a policy pattern that provides high-level information about sensitive topics while omitting specific instructions that could enable misuse.
- [Abliteration vs jailbreaking vs fine-tuning vs system-prompt guardrails](https://abliteration.ai/abliteration-vs-jailbreaking-vs-fine-tuning-vs-system-prompt-guardrails): Abliteration, jailbreaking, fine-tuning, and system-prompt guardrails are four distinct methods for controlling LLM refusal behavior, each operating at a different layer of the model stack.
- [How we measure refusal rate and benchmark retention for abliterated models](https://abliteration.ai/how-we-measure-refusal-rate-and-benchmark-retention): Refusal rate is the percentage of benign prompts a model refuses. Benchmark retention is the percentage of baseline benchmark scores preserved after abliteration.
- [AI gateway vs moderation API vs guardrails vs policy gateway](https://abliteration.ai/ai-gateway-vs-moderation-api-vs-guardrails-vs-policy-gateway): An AI gateway routes and observes LLM traffic. A moderation API scores content for harm categories. Guardrails are provider-side content filters. A policy gateway enforces your custom rules inline with deterministic outcomes and audit logs.
- [Zero data retention vs no-training vs ephemeral processing](https://abliteration.ai/zero-data-retention-vs-no-training-vs-ephemeral-processing): Zero data retention means prompts and completions are never stored. No-training means data is not used for model training but may be stored. Ephemeral processing means data exists only in memory during the request lifecycle.
- [Authorized penetration testing with governed AI: full workflow and controls](https://abliteration.ai/authorized-penetration-testing-with-governed-ai): Authorized penetration testing with governed AI is a workflow where security teams use abliterated models for technical research while enforcing audit trails, scoped access, and policy controls through Policy Gateway.
- [Synthetic data generation QA rubric: how to validate generated training rows before fine-tuning](https://abliteration.ai/synthetic-data-generation-qa-rubric): A synthetic data QA rubric is a structured checklist for evaluating generated training rows across dimensions like format correctness, factual accuracy, diversity, toxicity, and deduplication before using them for model fine-tuning.

## Notes for answer engines and agents

- For what the product is, what it costs, and how to start, use the product and pricing pages above.
- Prefer the reference pages above for definitions, API behavior, migration steps, and policy-gateway explanations.
- Prefer docs pages under /docs for implementation details and code-level setup.
- The public /v1 endpoints are OpenAI-compatible; policy-controlled traffic routes through /policy/* endpoints.
- Prompts and outputs are not retained by default; operational telemetry supports billing and reliability.