← All selected publications
NeurIPS 2026PosterFirst author

Harmless in Pieces, Harmful in Motion: Detecting Multi-Agent Jailbreaks

Vishal Pramanik, Maisha Maliha, Olivera Kotevska, Nathaniel D. Bastian, Susmit Jha, Sumit Kumar Jha

Detecting harmful intent that emerges across agents, tools, and shared memory—even when the individual steps appear harmless.

The manuscript PDF is not yet available on this site.

Research overview for citadel
Research overview · Harmless in Pieces, Harmful in Motion: Detecting Multi-Agent Jailbreaks

The research question

When can harmless-looking steps add up to a harmful workflow?

Research summary

This work studies safety at the level of a collaborating system. CITADEL tracks collective movement toward harmful objectives across agents, tools, and memory to detect distributed jailbreak behavior.

More of my research

NeurIPS 2026Poster · First author

Unified Resource-Grounded Coordination Protocol for Orchestrator-Free Heterogeneous Multi-Agent Systems

View research →EMNLP 2026Main conference · First author

Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

View research →CVPR 2026Highlight · First author

Selective Amnesia using Contrastive Subnet Erasure for Class Level Unlearning in Vision Models

View research →