Skip to content

Boundless Defensive Access

Abliterated models for authorized security testing

Investigate vulnerabilities and test your defenses with fewer model refusals.

  • 1K+ tokens per second
  • Zero data retention

Why abliterated

Standard models refuse requests that look like attacks

Testing a defense means reproducing the attack, which standard models often refuse. An abliterated model is modified to refuse less. It isn't more accurate, so its answers still need review.

What it's for

Find weaknesses before someone exploits them

  1. 01

    Code review

    Find and validate vulnerabilities before release.

  2. 02

    Defense testing

    Simulate misuse to see where your protections fail.

  3. 03

    Penetration testing

    Probe systems you're permitted to test, then check the fixes.

  4. 04

    AI application testing

    See how your AI app handles malicious inputs.

Pricing

Fast inference at a low price per token

DeepSeek V4.1 Flash 1017 Abliterated

Prices per 1M tokens

Input
$0.24
Cached input
$0.012
Output
$0.96
Tokens per second
1K+

Data handling

Zero data retention on every request

We never store, log or train on your prompts or outputs. We keep only what billing needs: token counts, model, timestamp and cost.

Apply

Apply for access

Tell us your company, the systems you plan to test, and your intended use. We review each application before granting access.

Model
Primary use case
How did you hear about us?
Agreements

Boundless handles your data under our Privacy Policy.

FAQ

Questions security teams ask before applying

Which models are available?

Defensive Access starts with DeepSeek V4.1 Flash 1017, abliterated, with a 1M-token context window. You can request another open-weight model in your application.

Who can apply, and how long does review take?

Companies doing authorized security work can apply, and we respond to every application within 24 hours.

It's a DeepSeek model. Does any data go to DeepSeek?

No, we host the model on our own infrastructure with zero data retention, so your requests never reach DeepSeek.

What is an abliterated model?

An abliterated model is an open-weight model edited to remove the internal direction that triggers refusals, so it rarely declines security requests that standard models refuse.

How does this compare with OpenAI's Daybreak?

Public benchmarks are promising: DeepSeek V4.1 Flash 1017 beats GPT-6 Sol with Daybreak access on one security task and ranks second only to it on another. Where we compete is scale, with over 1,000 tokens per second at roughly a tenth of GPT-6 Sol's list price per token.