Refract AI Collective is a vendor-neutral community helping public sector teams demystify, adopt, and secure artificial intelligence — bringing hackers, defenders, and builders together to stay ahead of a fast-moving revolution.
The institutions that serve the public — agencies, municipalities, schools, hospitals, and critical infrastructure — are adopting AI faster than they can secure it. The people defending them rarely have the budget, the vendor relationships, or the time to keep pace with a technology that reinvents itself every quarter.
Refract exists to close that gap. We are a non-commercial collective that pools knowledge, tooling, and expertise from across the security vendor community and channels it toward a single goal: helping public servants understand AI, deploy it responsibly, and defend it against real adversaries.
We translate frameworks into practice, run hands-on training and open working sessions, and create a place where the people protecting our constituents can learn in the open — together, and without a sales pitch attached.
AI security fails when offense, defense, and engineering work in isolation. Refract is built so all three sit at the same table.
Adversarial ML, red-teaming of LLMs, prompt injection, model extraction, and the attack techniques catalogued in MITRE ATLAS. We find the failure modes before adversaries do.
Monitoring, guardrails, governance, and incident response for AI systems. We turn frameworks like the NIST AI RMF into operational controls public sector teams can actually run.
Engineers and data scientists shipping AI into mission systems. We embed security by design — so the things we build for constituents are safe from the first line of code.
Practical, community-driven programs designed for the realities of public sector teams.
Plain-language briefings that cut through the hype — what generative AI, agents, and ML actually are, where they help, and where they quietly introduce risk.
Hands-on labs, workshops, and tabletop exercises on securing AI — from prompt-injection drills to building an AI governance program from scratch.
Threat modeling AI deployments against MITRE ATLAS, OWASP for LLMs, and the NIST AI RMF — mapped to controls that hold up under audit and attack.
We bring tooling, research, and experts from across the security vendor ecosystem to the public sector — neutrally, with no product agenda.
Everything maps back to one outcome: protecting the constituents and critical services that public institutions are entrusted to safeguard.
A standing community signal — sharing the latest techniques, incidents, and defenses so members never have to face a moving target alone.
The Collective organizes its work around open, authoritative frameworks — so members speak a common language with auditors, vendors, and adversaries alike.
The adversarial threat landscape for AI systems — tactics and techniques against machine learning, modeled like ATT&CK. Our north star for AI red-teaming.
The Risk Management Framework that helps organizations govern, map, measure, and manage AI risk — the backbone of our defender curriculum.
The most critical vulnerabilities in LLM applications — from prompt injection to insecure output handling. Required reading for our builders.
The foundational knowledge base of adversary behavior that ATLAS extends into the AI domain — the lingua franca of detection engineering.
Practitioners, public servants, researchers, and curious learners are all welcome. Membership is free — bring your color to the spectrum.