AI model safety testing for defense
Verify that the model you field behaves like the model you tested.
SichGate is an adversarial safety testing platform for small language models deployed in defense, national security, and edge environments. It runs entirely within your infrastructure, including fully air-gapped networks, and identifies where model safety degrades between training and deployment.
Supports
- GGUF
- GPTQ
- 4-bit quantization
- Any model size your hardware supports
- Air-gapped networks
Who SichGate serves
Defense technology companies
Building autonomous systems, unmanned aerial systems, and edge AI capabilities.
Prime contractors and systems integrators
Incorporating third-party or fine-tuned models into defense programs.
Test and evaluation teams
Responsible for verifying AI behavior prior to fielding.
Government and national security programs
Requiring independent assessment of models running on-premises or in disconnected environments.
The challenge
Models change before they reach the field.
AI models deployed at the edge rarely run in the form in which they were trained. They are fine-tuned for specific missions and then quantized, or compressed, to run on constrained hardware. Each transformation can alter how a model responds to adversarial or unsafe inputs.
Standard evaluations typically assess only the full-precision model. As a result, safety degradation introduced during compression often goes undetected until the model is already deployed. Aggregate performance scores can remain stable while specific safety behaviors deteriorate significantly.
FIG.1
| 1BaseFull precision, as trained | 2Fine-tunedAdapted to the mission | 3QuantizedCompressed for edge hardware. What actually ships. | |
|---|---|---|---|
| Standard evaluation | Tested | Sometimes tested | Not tested |
| SichGate | Tested | Tested | Tested |
SichGate closes this gap by testing the model at every stage of its lifecycle, including the exact configuration that will run in the field.
What SichGate evaluates
Four behaviors, tested at every stage.
Fail-safe behavior
Assesses whether the model declines out-of-scope, unauthorized, or unsafe actions rather than complying with them.
Adversarial robustness
Measures model resilience against manipulated, deceptive, and hostile inputs designed to induce unsafe behavior.
Quantization-aware safety drift
Compares base, fine-tuned, and quantized versions of the same model to identify where safety behavior degrades during compression.
Behavioral consistency
Verifies that the model maintains safe responses when inputs are incomplete, noisy, or degraded, as is common in operational conditions.
Each assessment produces a detailed report identifying failures by category and lifecycle stage.
Results map to SichGate's SG-1 through SG-4 attestation tiers, providing a consistent and auditable record of model safety.
Deployment options
Run it yourself, or with our team.
01 SichGate Air-Gapped
A licensed, self-installed package for secure and disconnected environments.
SichGate Air-Gapped is delivered as a self-contained package that your team installs and operates independently. It requires no external network connectivity and transmits no data to SichGate.
- Complete adversarial probe battery
- Quantization drift testing across base, fine-tuned, and quantized models
- Support for GGUF, GPTQ, and 4-bit quantization formats
- Locally generated reports and audit logs
- Installation and operating documentation
02 SichGate Managed Deployment
Supported installation and assessment for contracts and programs.
For engagements requiring direct support, SichGate installs and configures the platform within your environment and conducts the initial assessment in coordination with your team.
- On-site or secure remote installation
- Probe battery configured to your operational context and threat model
- Initial assessment conducted jointly with your team
- Findings review and remediation guidance
- A dedicated SichGate point of contact throughout the engagement
Need compliance evidence? See AI governance audits.
In development
Coverage beyond language models.
Organizations joining the waitlist receive early access and the opportunity to inform development priorities.
- In development
Vision models
Testing for perception, detection, and classification systems.
- In development
Multimodal models
Testing for models combining text, imagery, and sensor inputs.
- In development
Hardware-specific profiling
Evaluating models on the edge devices on which they are deployed.
Frequently asked questions
Do model weights or outputs leave our environment?
No. SichGate Air-Gapped operates entirely within your infrastructure. SichGate does not access your model weights, test data, or results.
Which models are currently supported?
SichGate supports small and mid-sized language models in base, fine-tuned, and quantized configurations. Because SichGate runs within your infrastructure, the model sizes it can test are determined by your available hardware. Vision and multimodal support is in development.
Does SichGate perform offensive testing or develop countermeasures?
No. SichGate focuses exclusively on safety assurance: verifying that models behave correctly, safely, and predictably. We do not develop capabilities to defeat or exploit operational systems.
Can SichGate support programs as a subcontractor?
Yes. SichGate can support defense programs as a subcontractor to prime contractors and systems integrators.
How do we get started?
Submit the contact form below with a brief description of your program and deployment environment. A member of the SichGate team will follow up.
Test what you field, not only what you trained.
SichGate provides defense and national security teams with independent, repeatable evidence that their AI models remain safe from development through deployment.