Investigate how AI can be manipulated.
How adversarial inputs, poisoned data, and compromised updates can change model behavior.
Four paths into a model's behaviour, and what Vraelis has on each today.
- Inputs at runtimeManipulated sensor or user inputAheadNo input checks or detector.
- Training dataPoisoned data before trainingAheadNo data provenance or poisoning tests.
- Model updateAn unauthorized or changed releaseWorkingRelease identity and approval checks in the experiment. They show a release changed, not where its bytes came from.
- Supply pathA wrong destination or changed componentsBuildingScripted fault tests specified, not run.
Model behaviour at runtime
Separate the attack from the conditions.
Which path does the attack use?
Sensor input, training data and model file each need their own threat model.
Where does a release check stop?
It shows a release changed, not model provenance: where the approved bytes came from.
What belongs in the baseline?
Lighting, viewpoint, normal variation and sensor faults, so ordinary conditions are not read as attacks.
Is it drift or an attack?
An anomaly is a reason to investigate, not proof of an attack.
What does a countermeasure cost?
Latency, memory and compute on the edge, plus misses and false alarms.
Research direction, with no product claim.
Nothing Vraelis has built tests, detects or blocks manipulated inputs or poisoned training data. No detector, no measured result. On the model-update path, the release identity and approval checks in the Vraelis Contour experiment show that a release changed, not where its bytes came from.
The one overlap is the supply path: unauthorized changes, a wrong destination or changed components, are specified as scripted fault tests, not run. That does not cover manipulated inputs or poisoned training data.
What is still open.
Which threats matter where
A vehicle and a control room face different attackers.
Adaptive testing
Fixed examples overstate a countermeasure.
Reproducible results
Model version, attacker access, input path and expected outcome.
One record for both
Release and input evidence in one ordered record.
Study these problems with us.
For researchers and academic teams.