Vorak

The Challenge

Most AI red-teaming tools rely on static jailbreak prompts and predefined attack libraries. They can verify known vulnerabilities, but they're poor at uncovering new ones.

I wanted to build something more intelligent—a framework that could adapt its attack strategy based on how a model responded, much like a human security researcher.

The goal was to create a system capable of discovering unknown failure modes through multi-turn adversarial conversations while producing reports that security and compliance teams could actually use.

The Solution & Technical Implementation

Vorak is an adaptive AI red-teaming framework built to continuously probe, evaluate, and escalate attacks against Large Language Models.

Instead of replaying fixed prompts, it generates increasingly sophisticated adversarial inputs until it either identifies a vulnerability or determines the model is sufficiently robust.

Adaptive Attack Engine

The core of the framework is an iterative evaluation loop:

1. Generate an adversarial prompt.

2. Evaluate the model's response.

3. Identify why the attack failed.

4. Generate a stronger, context-aware follow-up.

5. Repeat until success or a configurable evaluation threshold.

This allows Vorak to discover vulnerabilities that static prompt libraries would never reach.

Governance-First Reporting

Finding vulnerabilities is only half the problem.

Vorak automatically maps every security finding to enterprise governance frameworks including:

• NIST AI RMF

• MITRE ATLAS

• EU AI Act

• ISO/IEC 23894

This makes technical findings immediately useful for security and compliance teams.

Safe Security Sandbox

Generated code is never executed directly.

Instead, Vorak performs static analysis to detect:

• Dangerous imports

• Filesystem operations

• Network activity

• Potentially unsafe execution paths

This provides additional security without introducing unnecessary risk during testing.

Outcome

Vorak evolved into a reusable open-source framework for evaluating the safety of LLM-powered applications.

Highlights

• Adaptive multi-turn adversarial testing

• Automated attack escalation

• Governance-aware reporting

• Comparative evaluation across sessions

• Safe static code inspection

What I Learned

Building Vorak fundamentally changed how I think about AI security.

I learned that effective red-teaming isn't about collecting more jailbreak prompts—it's about creating systems that can reason, adapt, and continuously explore new attack paths. Just as importantly, I gained a deeper appreciation for the gap between identifying technical vulnerabilities and delivering security insights that organizations can act on.