Demos
Live interactive demos from our research on trustworthy AI systems.
GuardNet Live
A pre-inference graph-attention defense that detects both jailbreak and prompt-leakage adversarial prompts before they reach the target LLM. GuardNet constructs hybrid token graphs integrating sequential, syntactic, and attention-derived relations, then applies Graph Neural Networks for prompt-level detection and token-level localization — with no access to model weights.
Launch GuardNet Demo ↗DynaNoise Live
An adaptive inference-time defense against membership inference attacks. DynaNoise modulates injected noise based on per-query sensitivity estimated via Shannon entropy, preserving model utility while substantially reducing attack success rates — outperforming static defenses like DP-SGD and MemGuard across CIFAR-10, ImageNet-10, and SST-2.
Launch DynaNoise Demo ↗