Platform / Model Vulnerability Scan
Before it runs
Model Vulnerability Scan. Prove it is the model you approved.
Behavioral testing evaluates what a model does. Starseer examines what it learned: weights, activations, and internal structure, before it ever serves a request.
The problem
A backdoored model passes every benchmark you run.
Models arrive from repositories, vendors, and fine-tuning pipelines. Each hop is a step nobody attested, and the tests you run cannot see any of them.
Why testing does not catch it
01
The trigger is not in your test set
Trigger-conditioned behavior stays dormant until a specific input arrives. Until then the model scores exactly like a clean one, because behaviorally it is one.
02
Documentation is not evidence
A model card describes intended capability. It cannot tell you what the model actually learned, or what capability it picked up that nobody documented.
03
Every fine-tune resets the question
Quantizations, adapters, and community re-uploads each change the artifact. Approving a model once says nothing about the version running next quarter.
How it works
Continuous scanning, not a one-time check.
Every fine-tune, quantization, and vendor update re-enters the loop, so the approval never goes stale.
01
Baseline
Models from repositories, vendors, and fine-tuning pipelines establish the approved, known-good reference.
02
Fingerprint
Internal structure verified against that baseline: weights, activations, and architecture, not just outputs.
03
Analyze
Activation analysis and circuit tracing surface backdoors and trigger-conditioned behavior.
04
Attest
Provenance and integrity confirmed, and the verdict recorded as audit evidence.
Approved
Serves in production
Becomes the trusted baseline runtime measures against
Quarantine
Never serves
Forensic report generated for the deploying team
A request never reaches a quarantined model. An approved model becomes the reference for runtime detection.
What it catches
Six things a benchmark cannot tell you.
Each of these is read from the model's own internals rather than inferred from its answers.
Substitution and tampering
Verifies a deployed model matches its approved baseline down to internal structure, so a swapped or silently updated artifact is caught on arrival.
Backdoors
Surfaces trigger-conditioned behavior through activation analysis and circuit tracing, including triggers absent from every test you run.
Supply chain provenance
Confirms origin and integrity for models pulled from public repositories, vendors, and community re-uploads.
Undocumented capability
Identifies capabilities a model learned that its model card never mentions, which is where unexpected risk usually lives.
Fine-tune drift
Detects what changed inside a model after fine-tuning or quantization, not merely how its outputs shifted.
Missing audit evidence
Produces scan records aligned to MITRE ATLAS, NIST AI RMF, ISO 42001, and the EU AI Act as a by-product of every run.
In your pipeline
Runs where models already get approved.
The scan is a gate, not a report. It sits in the path a model already takes to production, so nothing new has to be remembered.
Gates your pipeline
Runs as a CI/CD deployment gate. Every result writes back as approve, quarantine, or investigate.
Feeds your runtime
Every model that passes the scan becomes the trusted baseline runtime detection measures against.
Deploys where you need it
SaaS, on-prem Kubernetes, or fully air-gapped, wherever your models live.
Covers
Open weights, fine-tunes, quantizations, community re-uploads
Integrates
CI/CD, model registries, SIEM and SOAR
Deploys
SaaS, on-prem Kubernetes, fully air-gapped
Evidence for
MITRE ATLAS, NIST AI RMF, ISO 42001, EU AI Act
Works with
Bring a model. We will tell you what is inside it.
A working session against a model you are about to deploy, or four minutes with the diagnostic.