AI Security & Interpretability Lab
AISy Lab studies how modern AI systems work on the inside, how they can be attacked, and how to make them safer and more transparent.
Research
We work at the intersection of interpretability and security: understanding models well enough to find their failures, and building defenses grounded in that understanding.
Mechanistic Interpretability
Reverse-engineering the internal circuits and representations of neural networks to explain what they compute and why.
Adversarial Robustness
Understanding jailbreaks, prompt injection and adversarial inputs, and building models and defenses that hold up under attack.
Privacy & Data Leakage
Measuring what models memorize and reveal about their training data, and methods that limit that leakage.
Auditing & Evaluation
Benchmarks, red-teaming protocols and auditing tools that make claims about model safety testable.
Current projects
All projects →LADA: Localization-Aware Distributed Alignment
Making LLM guardrails robust, auditable and maintainable by finding where safety behavior lives inside the model and making it harder to break.
GENESIS-AI
A reflexive generative AI ecosystem for cybersecurity that builds governance, compliance and adaptive protection into a single architecture.
Selected publications
All publications →-
NeuroStrike: Neuron-Level Attacks on Aligned LLMs
Network and Distributed System Security Symposium (NDSS), 2026
-
GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs
USENIX Security Symposium, 2026
-
Towards Backdoor Stealthiness in Model Parameter Space
ACM Conference on Computer and Communications Security (CCS), pp. 2863–2876, 2025
-
On the Vulnerability of Data Points Under Multiple Membership Inference Attacks and Target Models
IEEE Transactions on Dependable and Secure Computing 22(4), pp. 4022–4039, 2025