Research & notes
Latest paper · April 2026
Guardrails are memoryless. An adversary spreads one attack across sessions and slips past every detector. We show that bounded coreset memory is the architecture whose recall survives rewriting.
Two blockers kept security at the perimeter: massive activations, model-specific vectors. Platonic codes compress routes into a shared frame so behavioral anomaly detection can sit next to inference, including zero-day prompt injection attacks.
A unifying picture for ABA: anomaly detection across event, intent, and residual-stream space — from pub/sub tool calls to belief-state geometry inside the model.
Agent identity is not a static concept. It's a trajectory through an information manifold — and every threat is a deviation from the intended path.
Intent, identity, threat, and policy as four faces of a single geometric object — a unified frame for how agent security should be measured.
Exploring adaptive prompt injections and how reasoning-layer security provides superior protection against evolving AI threats.
Understanding how atomic attack components combine to form multi-stage threats, and how token-level analysis reveals complex attack patterns before they execute.
As agentic AI systems take actions, security shifts from 'what happened?' to 'what was the model trying to do — and why?' Thought-level entities must become first-class in security monitoring.
Design partners
You bring a workload; we sit with your team until it proves itself, or doesn't.
Become a design partner