MediumAdversarial Evasion◐Demonstrated
Adversarial / evasion inputs
Security & RobustnessDescription
Crafted perturbations cause misclassification or unsafe handling (e.g., evading a fraud/abuse classifier).
Example scenario
Obfuscated text evades the toxicity/fraud filter on a complaints channel.
Real-world evidence◐Demonstrated
BadNets and subsequent academic work rigorously demonstrated backdoor injection via poisoned training data in controlled settings. No confirmed production-scale poisoning attack against a deployed BFSI AI system has been publicly disclosed.
Primary mitigations
- Adversarial training
- input normalisation
- ensemble detection
- perturbation testing
- anomaly detection.
Detection signals
Adversarial-robustness benchmarks; evasion-attempt monitoring.
Mitigating controls
4 Non-agentic controls