Red teaming finds where a system breaks; blue teaming makes sure it holds. We turn the findings into defenses a defender can actually run — hardened protocols, positive-security guardrails, and provenance artifacts — and we build to the boundary the red-team work proved, so no control ships a guarantee it cannot deliver.
Our blue-teaming methods come out of published, peer-visible research — not a proprietary black box. Each technique below is backed by a paper you can read.
You cannot ship a defense that provably cannot hold. The Defense Trilemma establishes the boundary where prompt-injection wrapper defenses must fail — continuity, utility preservation, and complete safety cannot coexist. We build to the boundary, not past it, so a control does not fail silently in production.
Read the paper: The Defense Trilemma →Agentic systems are only as safe as the tools they call. ETDI hardens the Model Context Protocol against tool squatting and rug-pull attacks with OAuth-enhanced tool signatures and policy-based access control, so a malicious or mutated tool cannot be silently substituted into an agent's workflow.
Read the paper: ETDI →Detecting every bad input is impossible; allowing only the right input is tractable. QFIRE is an inline prompt firewall for healthcare AI agents that combines positive-security scope constraints with an asynchronous detector graph — 0.83 recall where generic injection classifiers reach only 0.40.
Read the paper: QFIRE →An agent's actions must be reconstructable after the fact. MAIF embeds provenance, access control, and audit evidence into the data objects agents use, so each artifact carries the signed chain of who touched it, what it was allowed to do, and where it came from.
Read the paper: MAIF →A defense earns its place when it runs. Our released technology includes a deployed detection model (PhishSense, with a locally runnable quantized build) and a coordinated vulnerability disclosure — CVE-2025-48913 in Apache CXF — that led the maintainer to restrict the dangerous protocol. The same loop runs through the blue-teaming work: prove the boundary, build the control, ship the artifact, disclose the finding.
The red team surfaces where a system fails; the blue team makes sure it holds. We run them as a single engagement — the same evidence informs both sides, so a fix is tested against the failure that motivated it, not a separate, easier problem.
We take on a small number of defensive engagements — hardening agentic protocols, building positive-security guardrails, and shipping provenance artifacts. Tell us what control you need to hold.
Request a project →