VIRIDISSECURITY
Request assessment

AI agent and MCP security

Security assessmentsfor AI agents.

Understand who can act, what can fail, and how your system recovers. Viridis reviews AI agents and MCP systems, then helps you choose controls with evidence your team can inspect.

Request an assessmentExplore the Maxwell demo
Inspect the research · Open preprint10.5281/zenodo.21855705
ASSESSMENT / 01ADVISORY
MANDATORYVIABLE?HOLD ON UNKNOWN
AUTHORITYSAFETYRECOVERYEVIDENCE
DisturbanceDeclared
UncertaintyHold
ExecutionDisabled
AUTHORITY MAPSCONTROL BOUNDARIESRECOVERY SCENARIOSEVIDENCE & LIMITSPRACTICAL DEFENSES

What customers buy

Understand the risks. Choose the controls.

Start with the system you own and the decision you need to make. Assessments map authority, investigate control gaps, and compare containment and recovery under agreed scenarios.

01

Agent resilience assessment

Compare two agent or MCP architectures under declared compromise, outage and recovery conditions before committing to the faster design.

View offering
02

Authority and architecture review

Map authority, dependencies, fallback paths and proof obligations to the exact system boundary your team owns.

View offering
03

AI control-boundary review

Test tool, retrieval, memory and action boundaries against instruction injection and unsafe authority transfer.

View offering
04

Vulnerability evidence review

Move from source-grounded hypothesis to reproducible evidence without promoting scanner output into a security claim.

View offering

Featured product · Open source

Maxwell's Defense

A proof-of-work gate for endpoints you operate. The reference middleware accepts a valid challenge once; another accepted request requires a valid solution to a fresh challenge.

Explore the illustrated flow, inspect the code, and discuss whether this control fits your application. Shared state and capacity planning are part of the deployment decision.

Explore the demoSource and deployment guide

One challenge · One acceptance

  1. 01 Receive a challenge
  2. 02 Solve and submit work
  3. 03 Accept once
  4. 04 Reject replay · Issue fresh challenge
Illustration of the reference middleware flow.

The assessment process

From system boundary to practical decision.

The kernel orders authority, safety and mandatory viability ahead of nominal performance. Unknown mandatory evidence produces HOLD.

  1. 01

    Bind authority

    Name the owner, system boundary, protected obligations and allowed evidence.

  2. 02

    Map dependencies

    Trace agents, tools, providers and recovery paths to their shared roots.

  3. 03

    Apply disturbances

    Compare mandatory viability, containment and recovery before nominal speed.

  4. 04

    Issue the receipt

    Preserve inputs, uncertainty, limitations and the non-actuating decision.

Research translated into product

Methods and research with visible limits.

Static scans return indicators to investigate; a clean result does not prove absence of vulnerabilities. Assessment recommendations cite their evidence and limitations. Research statements keep their assumptions separate from a customer's deployed system.

Open the research and proof atlas
DECISION RECEIPTED25519 / HASH-BOUND
Architecture boundaryDeclared
Mandatory uncertaintyFails closed
Deployment authorityNot granted
input_sha256 → decision_sha256 → signature

Design-partner assessment

Bring two architectures and one decision that matters.

We return the authority map, disturbance comparison, common-mode analysis, limitations and signed advisory receipt.

Request a resilience assessment