VLM Security
Multimodal attacks, physical prompt injection, scene manipulation, grounding, and adversarial perception.
Independent AI security research / 01—04
Where a signal becomes an action.
Independent research on vision-language models, Vision-Language-Action systems, robots, autonomous agents, and the persistent state that connects them.
Current focusVLM, VLA, and robotics security.
Security boundaryFrom perception to physical action.
MethodAdversarial and reproducible experiments.
FoundationAgents, memory, provenance, and shared state.
Working paper
View papers →Working paper preview · September 2026
StaleAction
Physical TOCTOU Attacks Against Vision-Language-Action Robots
A threat model for post-observation physical state changes that exploit the gap between a robot’s perception and action.
Research areas
Our research program →VLA & Robot Security
Action freshness, physical TOCTOU, runtime validation, robotics, and perception-to-action failures.
Agent & Multi-Agent Security
Tool use, authority boundaries, coordination, shared state, adversarial propagation, and emergent behavior.
Memory & Provenance
Retrieval, poisoning, lineage, correlated evidence, persistence, and cross-agent propagation.
One continuous question
From persistent memory to physical agency.
NEUROSIS began by asking what happens when autonomous agents accumulate, retrieve, modify, and share knowledge over time. We now follow that question into systems that can perceive and change the physical world.
MemoryAgentsMulti-agent systemsPerceptionEmbodimentAction
The original public memory experiment remains active as part of this larger research program.
Read our story