Skip to main content

Air-gapped AI for cybersecurity. Inside your wire.

Excalibur runs CyberAGI's own models on NVIDIA DGX Spark hardware in your datacenter, with no external AI API. It maps, tests and reports on your environment with zero data egress.

BOUNDARY · excalibur-01

Live
Data egress
0bytes out
Models
Localon device
Inbound feeds
2you control
Approval
1pending
  1. ▸ boundary check
  2. · uplink: none · external AI API: none
  3. · models: CyberAGI private, on-device
  4. ▸ threat intel sync · inbound only
  5. ✓ EPSS scores imported by @ops
  6. · NVD update staged · awaiting operator
  7. ▸ agents at work
  8. ✓ attack surface mapped · results stored locally
  9. ✓ threat model updated against your architecture
  10. ⊘ BLOCKED outbound request · boundary held
  11. ⧗ requesting approval @sarah.k

Product interface shown with illustrative sample data.

Every sealed network runs into the same question. “Which AI security tool can do real work on a network that never connects out?”

On-prem is not the same as air-gapped.

Plenty of AI security tools install inside your network and still reach out for their model. In a sealed environment, that one call is the deal-breaker.

Cloud AI

Your findings, prompts and asset data travel to a vendor's servers to be processed.

On-premises AI

The application sits on your hardware, but it may still call an outside AI API, a cloud control plane or a license server.

Air-gapped AI

Models, inference, data and threat intelligence all run inside an isolated network. Pull the uplink and the work continues.

What air-gapped has to mean in practice.

A demo with the network cable pulled is the real test. Here is how Excalibur is built for it, next to how AI security tools usually work.

Hardware-only delivery: one box, one purchase order, one boundary.

CVE corpus, nuclei templates and TTP playbooks pre-loaded at delivery.

You control inbound sync from NVD and EPSS. Nothing flows out.

CapabilityExcaliburTypical AI security
Where models runNVIDIA DGX Spark in your datacenterVendor cloud
Model dependencyCyberAGI's own models, no external AI APIOften a frontier model behind an API
Customer data egressZeroCloud or hybrid processing
Authorization boundaryRuns inside your existing ATOA new external cloud service
Model evolutionPrivate fine-tuning on your hardwareNone

The whole security program, inside the boundary.

“Which tools come off my invoice on day one, and which are you only sitting on top of?” Excalibur runs attack surface mapping, threat modeling, pentesting, correlation and compliance on one box in your network.

Discover

Attack Mapper

Continuous external attack surface mapping with a native scanner. No third-party cloud dependency, no asset data leaving your environment.

Model

Threat Modeling

STRIDE, ATT&CK and kill-chain modeling against your architecture, mapped to your CUI flows and classification boundaries.

Validate

Agentic Pentesting

Atomic Red Team tests, MITRE ATT&CK simulations and API testing inside a scope you set, human-verified before any action.

Correlate

Exposure Fabric

Direct integrations for Wiz, CrowdStrike, Qualys, Defender and Tenable. One-way ingestion, never egress.

Govern

Compliance

Control questions are generated and filled from lifecycle data, and findings map to CMMC, NIST 800-171 and PCI-DSS controls.

Learn

Self-Evolving AI

Models fine-tune on your private data, on your hardware, and get sharper every cycle.

Sealed from the outside. Answerable to you.

Inside the boundary, agents do the repetitive work. Anything critical stops and waits for a named person on your team.

  • You set the scope. The agents hold every boundary in it.

  • Critical findings wait for approve, reject or request info.

  • Every step is logged on your hardware, where your team can review it.

Approved@sarah.k

Probe: SQLi on /api/users

Boundary heldscope-v4.2 §3.1

Attempt to read production PII

Awaiting approvalcritical · CVSS 9.8

Submit finding: SQLi → SSRF → IAM

“I was initially skeptical about AI's impact on the cybersecurity industry, until I started working with Excalibur.”

Excalibur’s modules and technology have significantly improved my ability to manage multiple security frameworks and architectures. The Posture model delivers actionable insights that enhance the effectiveness of Tier 1 analysts while providing an excellent learning path, reducing the time I need to spend on foundational training. It also does an outstanding job of modeling potential threats and identifying security concerns.

The recently introduced Govern model is ahead of the curve when it comes to reducing the time required for compliance management and audit preparation. The ability to answer framework-specific questions simply by uploading policy documents is a tremendous time saver. Our previous platform, Apptega, which remains a solid solution in terms of cost and capabilities, could not extract answers directly from documentation, making this feature a true game changer.

I firmly believe AI should be used to make security professionals more effective and efficient, freeing them to focus on meaningful security work rather than replacing jobs. Excalibur fully embraces that philosophy and continues to deliver rapid innovation and improvement.

I’m excited to see where this product goes next.

Loial (A Logicalis Company)

Cody Jackson

Principal Engineer, Loial (A Logicalis Company)

Questions buyers ask about air-gapped AI.

Air-gapped AI runs entirely on hardware inside an isolated network with no outside connection. The models, the inference, your data and the threat intelligence it works from all stay inside. Excalibur is built this way for security work: it runs on NVIDIA DGX Spark hardware in your datacenter.

On-prem AI runs on your hardware but can still depend on an outside AI API, a cloud control plane or vendor updates. Air-gapped AI keeps working with every outside connection removed. Excalibur runs CyberAGI's own models with no external AI API, so it can do both.

No. Excalibur runs CyberAGI's own private models on the hardware in your datacenter. Enterprise customer data is not sent to, stored in or processed by CyberAGI-operated systems.

The CVE corpus, nuclei templates and TTP playbooks are pre-loaded at delivery. After that, you control inbound sync from NVD and EPSS. Data comes in when you allow it. Nothing flows out.

CyberAGI is not FedRAMP Authorized and does not claim to be. FedRAMP covers vendor-operated cloud services, and Excalibur is software deployed inside your own environment. It runs within your existing authorization boundary, under your Authorizing Official and your ATO process.

The Govern stage covers CMMC L1 and L2, NIST 800-171 and PCI-DSS, and NIST SP 800-53 content ships with the platform. More frameworks can be added on request. Excalibur helps your team meet these obligations; deploying it does not make you compliant on its own.

Defense customers go from purchase order to running platform in the same week.

Nine Teams. One closed loop.

Discover. Model. Validate. Correlate. Posture. Remediate. Govern. Predict. Learn. Each stage sharpens the next. See what happens when they close the loop.

Already know what you need? Configure Excalibur