Model APIs for High-Risk Work | abliteration.ai

Abliteration

2 min read Original article ↗

Platform

Models, policies, and data.

Access models for security and safety work, apply your organization’s policies, and generate data for training and evaluation.

Model API

Connect the tools you already use.

Access models for security and safety work that mainstream AI often refuses. Keep your OpenAI or Anthropic client; change the endpoint, API key, and model.

Python example using the OpenAI SDK
import osfrom openai import OpenAI client = OpenAI(    base_url="https://api.abliteration.ai/v1",    api_key=os.environ["ABLITERATION_API_KEY"],)response = client.chat.completions.create(    model="abliterated-model",    messages=[{"role": "user", "content": "Explain the security risks of Invoke-Expression."}],)print(response.choices[0].message.content)

Read the API documentation

Policy Gateway

You decide what your team can ask for.

Write rules for the work your organization authorizes. Requests outside them are refused before they reach the model.

Explore Policy Gateway

A blue request path passes through a sand-colored policy gate. A second, sand-colored request path stops at the gate.

Requests inside your policy pass through the gateway to the model. Requests outside it stop at the gateway.

Training Data

Turn one brief into training and evaluation data.

Describe the cases your classifier or model must handle. You receive records in the formats you train and test with.

Example brief

Payroll phishing and legitimate payroll messages for an email classifier, with hard cases held out for evaluation.

  • Labeled examples

    Train a classifier

    Your payroll account is locked. Confirm your password at payro11-update.example within 2 hours.

    Phishing

    October payslips are available in the employee portal. Sign in using your saved bookmark.

    Legitimate

  • Preference pairs

    Tune a model’s responses

    Is this phishing? “Your deposit is on hold. Reply with your password to release this month’s pay.”

    Preferred
    Phishing. It asks for a password by reply, which a payroll team would not request.

    Rejected
    Legitimate. Payroll notices about held deposits are often urgent.

  • Evaluation sets

    Measure a model before release

    Your W-2 is ready. Download it at payro1l-docs.example/secure.

    Expected: PhishingLookalike domain

    Open enrollment closes Friday. Review your benefits in Workday.

    Expected: LegitimateUrgent but safe

Explore Training Data

Evaluate it on your own work.

Your prompts and outputs aren’t stored by default or used to train models. Review our evaluations and data policies, or talk with us about your organization’s requirements.