Boundless officially launches Defensive Access: providing authorized teams with dedicated AI models to help defenders counter the surge in attacks.

Source: Boundless

AI inference service provider Boundless today announced the launch of Boundless Defensive Access, a program that provides vetted security teams with access to Abliterated (low-refusal) models fine-tuned for security research. The first model available is a security research version of the open-source model DeepSeek V4.1 Flash 1017. The program aims to solve a long-standing challenge for defenders: general-purpose AI models often refuse legitimate security testing requests that resemble attack patterns, while attackers face no such constraints.

Defenders Should Not Fall Behind Attackers in Tools

The security industry is facing an unprecedented volume of attacks. Penetration testing, vulnerability verification, and attack simulation inherently require defenders to think like attackers. However, such requests closely resemble real attacks in form, and even when security teams are testing systems they own and are authorized to assess, general-purpose models frequently refuse to assist.

This has led to a situation where malicious actors can deploy and modify open-source models on their own, while rule-abiding defensive teams are at a disadvantage at the tooling level. The goal of Defensive Access is to put equivalent AI capabilities into the hands of defenders under strict admission requirements.

First Model: DeepSeek V4.1 Flash 1017

Early benchmark results released by independent evaluation firm Vals AI on October 20 show that the standard version of DeepSeek V4.1 Flash 1017 outperforms GPT-6 Sol on its multi-step cybersecurity tasks built on simulated enterprise networks. Boundless offers a security research version of this model that reduces unnecessary refusals in legitimate security testing scenarios.

To meet the demands of high-precision, real-time security tasks, Boundless provides industry-leading inference performance support:

Over 300+ Tokens/sec: Delivers ultra-fast response for security automation agents and real-time penetration testing tools.

Top throughput: Measured at up to 349.90 tokens/sec on Vercel AI Gateway, ranking first in the industry.

Enterprise-grade hosting: Eliminates the burden of security teams deploying and maintaining large models themselves, providing 24/7 stable operation assurance.

"Traditional models' safety restrictions often backfire, ironically blocking those who genuinely want to implement protections," said the Boundless team. "Through Defensive Access, we are equipping defenders with unrestricted frontier model capabilities powered by an ultra-fast, high-throughput inference engine."

Four Core Application Scenarios and Application Mechanism

Boundless Defensive Access focuses on four defensive scenarios:

Code Review: Identifying and verifying potential security vulnerabilities before systems go live.

Defense Testing: Simulating real hacker tactics to identify blind spots in existing defense networks.

Penetration Testing: Conducting in-depth probing of systems within authorized scope and verifying remediation solutions.

AI Application Testing: Testing AI applications' defensive capabilities against malicious inputs and jailbreak instructions.

To ensure the technology is used safely and responsibly, Defensive Access employs a strict application review process. Every applicant team must undergo compliance qualification screening to ensure models are used solely for compliant security research, red team exercises, and penetration testing. Additionally, for enterprises with special requirements, the Boundless team also supports customized deployment of other open-source Abliterated models.

How to Apply

Boundless Defensive Access is officially open for applications starting today. Security practitioners and enterprise teams can visit the official website to submit applications.

About Boundless

Boundless is a high-performance AI inference infrastructure platform focused on open-source models, dedicated to providing ultra-low-latency, high-throughput model hosting and inference services for AI-native enterprises and developer teams.

This article is contributed content and does not represent BlockBeats' views.