Introducing Boundless Defensive Access: abliterated models for authorized security testing
Last updated: October 23, 2026
Contents
Security teams need to understand how their systems can fail before someone exploits them. That work can involve examining malicious code, simulating attacks, and testing how applications respond to misuse.
AI models can help with this work. But they can also refuse legitimate security requests because those requests resemble harmful ones.
We’re opening applications for Boundless Defensive Access, starting with an abliterated variant of DeepSeek V4.1 Flash 1017. Apply here.
What is an abliterated model?
An abliterated model is a language model modified to reduce its tendency to refuse requests.
The modification targets internal patterns associated with refusal. Researchers have shown that changing these patterns can make a model answer requests it would otherwise reject.
Reducing refusals does not make a model more accurate or a better security expert. It changes which requests the model is willing to engage with. Its answers still need to be tested and reviewed.
How can that help defenders?
Defensive testing often requires reproducing the behavior you want to prevent. To check whether an application resists an attack, you need to explore how that attack could work.
Abliterated models can support that investigation within an authorized testing environment.
- Review code before release. Investigate potential vulnerabilities, examine suspicious behavior, and develop tests to check whether a finding is valid.
- Test application defenses. Simulate misuse in a controlled environment to assess how your application responds and where its protections need improvement.
- Support penetration testing. Explore weaknesses in systems you have permission to test, reason through findings, and check whether proposed fixes address the issue.
- Evaluate AI applications. Develop adversarial test cases to assess how an application handles malicious inputs and attempts to redirect its behavior.
The goal is to find weaknesses, validate them, and address them before deployment.
Why DeepSeek V4.1 Flash 1017?
Early benchmarks from Vals AI show promising results on multi-step cybersecurity tasks, making it a strong candidate for security teams to evaluate. If you need a different model, tell us which model you want and what you’re testing. Our team can work with you to deploy an abliterated variant for your use case.
Why run it on Boundless?
Security testing involves repeated analysis, test generation, and follow-up. Time spent waiting for each output adds up.
Boundless serves the abliterated variant of DeepSeek V4.1 Flash 1017 at 300+ tokens per second. Faster generation means less time waiting for long analyses and test cases, leaving more time to investigate findings and validate fixes.
We handle model hosting and serving, so your team can focus on testing its defenses.
How does access work?
Boundless Defensive Access is intended for authorized security research, red teaming, and penetration testing.
Apply with your company, the systems you plan to test, and your intended use. We review each application before granting access.
Apply for access
We’re inviting teams working on security research, red teaming, and penetration testing to apply.
Starting with DeepSeek V4.1 Flash 1017, abliterated.