Security teams drown in findings.
CyberGuard sorts and prioritises findings so analysts start with what matters.
Security triage and threat analysis on a small model.
CyberGuard is a small specialized model for security triage, threat analysis and security reasoning — sized to run on a single 24GB GPU, inside your own perimeter.
Why it exists
Problems we are solving
Security data should not leave the building.
CyberGuard is small enough to run on one local GPU, so logs and code stay inside your perimeter.
False positives are what teams feel first.
CyberGuard is measured on false-positive rate on safe code, not only on what it catches.
How it is measured
Vulnerability-detection precision, vulnerability-detection recall, and false-positive rate on safe code. The production bar is false positives under 10% and false negatives under 10% — following the AWS Deception Benchmark framing that false positives are the operational number security teams feel first.
Benchmarks
Cybench, CyberSecEval and the AWS Deception Benchmark — public suites, so results can be compared and reproduced.
What it does
Built for security work.
Security triage
Sorts and prioritises security findings by severity and context.
Threat analysis
Reasons over threat and attack data in the security domain.
Local by design
Sized for a workstation or a single 24GB GPU. Nothing has to leave your network.
Served by Ember
Runs behind Ember, Aitium's inference engine, through an OpenAI-compatible API.
Process
How it is built
Measure
Precision, recall and false-positive rate on public security benchmarks.
Publish
Results are published in full, including the misses.
Ship
Weights and a model card, served through Ember on your own hardware.
Access
Follow CyberGuard
CyberGuard is in evaluation against public security benchmarks. Follow it to get the results and early access first.
Follow CyberGuard- On release
Weights
Model weights and card
Requirements
- A 24GB GPU or workstation for local runs
- Ember for serving
Questions / answers
CyberGuard FAQ
What does CyberGuard do?
CyberGuard is a small model for cybersecurity triage, analysis, and domain reasoning.
Where does my data go?
Nowhere. CyberGuard runs on your own GPU or workstation, so logs and code stay inside your perimeter.
How is it measured?
Precision, recall and false-positive rate on safe code, using Cybench, CyberSecEval and the AWS Deception Benchmark.
How will it be served?
Through Ember, Aitium's inference engine, behind an OpenAI-compatible API.
Models are served through Ember, on hardware we control.