AI systems are not invulnerable. As they spread, so do the attack vectors – and awareness has to keep pace.
Prompt injection
With prompt injection, attackers try to override the instructions given to an AI system through cleverly worded input. That can lead to confidential system information being disclosed or unwanted actions being carried out.
Data poisoning
If attackers can influence a model's training data, they can manipulate its behaviour deliberately. KOSMO relies on on-premise operation and RAG – which makes data poisoning considerably harder.
Model inversion
Through targeted queries, attackers can try to draw conclusions about training data. Here too, local operation and not passing data on both help.
How does KOSMO protect against this?
- Role-based access rights prevent unauthorised access to data
- No internet access for the model without explicit clearance
- All queries and answers can be logged
- On-premise operation minimises the attack surface
