We test AI,
and everything it runs on.

Tech industry giants have trusted us with their most security-critical testing since 2002. We're curious people who love this work.

From AI models to firmware

We test the models, agents, and AI features you're shipping, plus the apps, cloud, and devices they run on. Turns out two decades of testing the full stack is great preparation for AI.

Nemesis, our AI-native offensive security platform

A scanner flags what might be wrong. Nemesis proves what's real. Our consultants scope the engagement, direct the work, and monitor it as it runs. High-integrity agents do the investigating. We guarantee they operate as intended and produce high-quality results: they reason across your code and live systems, chase down what matters, and prove what they find against the live target. Nothing reaches you until a consultant has signed off.

nemesis > investigate --engagement acme-webapp
[agent] tracing auth flow across 6 modules via call graph + semantic search
[agent] candidate: session token still accepted after logout
[net] POST /api/orders (forged JWT) -> 200 OK
[found] authentication bypass confirmed against live target
[agent] chaining bypass with IDOR on /orders/:id -> cross-tenant access
[+] 3 critical findings validated - report draft ready for review
[human] consultant signing off before delivery

Microsoft's model card for MAI Code 2 describes our offensive cyber evaluation

Microsoft's model card for MAI Code 2 and MAI Code 2 Flash describes Casaba as "an independent assessment firm selected for their expertise in cybersecurity and experience evaluating advanced AI models and systems for cyberrisk." We assessed both models for offensive cyber capability across eight evaluation areas. Microsoft published what we found, including that both models remain susceptible to some image-based and multi-turn jailbreaks.

Read what Microsoft published

Red teaming M365 Copilot since 2024

Microsoft brings us in to test Copilot and web applications across the M365 ecosystem. We use the OWASP Top 10 for LLM Applications as a baseline and test Responsible AI harm categories alongside security. The reports are public on Microsoft's Service Trust Portal.

Read the Copilot case study
20+
Years in business
60+
Security professionals
3
Global regions
CREST
Approved provider

Microsoft, Amazon, Meta, Adobe, Costco, NetApp, GE, and hundreds more.

Let's talk

Tell us what you're working on. You'll talk directly to the people you'd work with.

Get in touch