About us

We think every deployed model deserves a threat model.

Ekushield was built by machine learning and security engineers who kept seeing the same gap: teams rigorously test model accuracy, then ship it to production with no idea how it behaves under adversarial input.

Why we exist

Adversarial machine learning research has produced well-understood attacks — FGSM, PGD, boundary attacks, data poisoning — for close to a decade. Most of that work stays inside research papers. Ekushield's job is to turn it into a report a product or security team can act on in an afternoon, not a PhD thesis.

We don't believe in a single "robustness score" that hides its own math. Every number Ekushield reports — clean accuracy, critical epsilon, hardening priority — traces back to a documented method you can read on the How it works page.

How we work

Small team, model-agnostic by design. We support tabular classifiers today, with image and LLM guardrail testing available through the API, and we prioritize new attack coverage based on what customers are actually defending against — not what's trending in the literature.

Where possible, computation happens client-side or inside your own infrastructure. We would rather ship you a report than ask for your training data.

Principles

Show the math

Every score has a documented, reproducible method behind it.

Default to your infrastructure

We minimize the data you have to hand us to get a useful answer.

Ship fixes, not just findings

A vulnerability report without a ranked remediation plan isn't finished.

Want to work on this with us?

We're a small, remote-friendly team hiring for ML security research and platform engineering.

Get in touch