Some of the most useful questions in cybersecurity are also the ones general-purpose AI models tend to refuse to answer.
Ask a model to investigate an authentication bypass, analyze suspicious code, or explore how a vulnerability could be exploited, and you might get a refusal instead of useful analysis.
Those restrictions make sense for general-purpose AI, but they can also get in the way of legitimate security research and defense.
That's why we're introducing two new programs at LunaRoute: Cybersecurity and Cyberdefense.
Two sides of the same problem
Cybersecurity focuses on authorized vulnerability research, testing, and remediation.
Imagine a security team using AI agents to analyze a codebase, explore potential vulnerabilities, test them in controlled environments, and validate proposed fixes. Or a security consultancy running multiple investigations in parallel, looking for weaknesses that might otherwise go unnoticed.
Cyberdefense focuses on understanding threats, building detections, and responding to incidents.
A security operations team could use AI to analyze suspicious scripts, investigate alerts, develop detection rules, and test those rules against variations of an attack. This becomes particularly interesting as attackers themselves increasingly use AI to automate and adapt their techniques.
In both cases, the models assist security professionals rather than replace their judgment.
More capable models, controlled access
We're planning to offer abliterated model variants, including GLM 5.3 Cyber and DeepSeek 4.1 Flash Cyber, with reduced refusal behavior for legitimate security work.
But fewer refusals don't mean unrestricted access.
Both programs will require identity verification (KYC/KYB), approval, and separate terms. Unlike our standard Zero Data Retention offering, prompts and responses will be retained for 30 days by default for security purposes, with ZDR available through additional approval.
We believe verified access is a better approach than simply preventing models from engaging with difficult security questions.
Exploration is the whole point
Both security research and defense require a lot of experimentation. You investigate a hypothesis, test it, discover something unexpected, and try again. You might explore hundreds of potential vulnerabilities or test dozens of detection strategies before finding something useful.
This is also where LunaRoute's approach to inference becomes particularly interesting.
When you don't have to think about how many tokens each experiment consumes, you can push those explorations further. Run more tests, investigate more possibilities, try approaches that might fail, and discover things you wouldn't have otherwise.
That's the kind of freedom we want to bring to security teams.
Join the early-access waitlist
Both programs are currently in planning, and we're opening waitlists for researchers, security teams, and organizations interested in these capabilities.
If this sounds like something your team could use, we'd love to hear from you.