FIELD NOTES
Research for
working Defenders.
Short, technical, and grounded in operational reality.
FIND THE OPERATIONAL NOTE
Filter by the problem you are working.
Detecting unauthorized agent delegation and task propagation
Trace parent-child agents, delegated authority, task queues, containment decisions, and confirmed downstream consequences.
When retrieval becomes the attack path
Incident response for poisoned RAG pipelines: trace contamination, contain retrieval, preserve evidence, and rebuild trusted indexes.
When your AI evidence cannot be trusted
Preserve execution paths, provenance, prompts, retrieval, memory, tool actions, and independently verifiable AI evidence.
Incident response for compromised machine identities
A technical response guide for containing service accounts, workload identities, API keys, OAuth applications, certificates, delegated tokens, and their persistence paths.
The first 60 minutes of an AI security incident
A commander-and-operator timeline for stopping autonomous activity, revoking unsafe authority, preserving evidence, and bounding downstream impact.
Label before retrieval: securing data before AI touches it
A control architecture for preserving ownership, classification, authorization, retention, and tenant boundaries through RAG and agent memory.
Guardrails that can say no: OPA policy enforcement for AI agents
A practical architecture for independent policy decisions at agent tool boundaries, with approval binding, failure behavior, and audit requirements.
Incident response for autonomous AI agents
A practical framework for scoping identity, memory, tools, and delegated actions when an AI workflow goes wrong.
What to log before your first AI incident
The minimum viable telemetry model for prompts, retrieval, policy decisions, tool execution, and human approvals.
Ransomware containment under pressure
A technical decision framework for isolation, control-plane protection, evidence preservation, and staged recovery.