Security of AI Training Interconnects
Threat modeling the high-speed networks that move data between accelerators in distributed AI training clusters.
Read moreLoading
Preparing the next field note…
Loading
Preparing the next field note…
Security engineering leader and researcher · Bellevue / Seattle
I’m a security engineering leader with a deep technical background, now working at the intersection of AI, infrastructure, and cybersecurity.
I led offensive security work at Amazon and built the information security function for mainland-China iCloud operations from pre-launch infrastructure through mature 24/7 operations. I now apply that experience to AI infrastructure and agent security.
Current direction
My active work looks at the security of distributed AI infrastructure and long-term agent memory. These projects draw on familiar security disciplines: threat modeling, adversarial testing, detection, isolation, and incident-informed design.
I’m also exploring cyber-capability evaluation, adversarial AI, and compute verification. In compute verification, I see a parallel with network security monitoring: turning activity inside a complex system into reliable, security-relevant evidence.
Why AI security
Mechanistic interpretability first drew me into AI safety because it felt like reverse engineering: understand the system internally, not only through its outputs. I still care about that long-term promise, but I have become more realistic about what current methods can establish.
My present focus reflects comparative advantage: bringing offensive and defensive security experience to AI systems already being built while continuing to deepen my machine learning and computer science foundations.
My professional journeySelected work
Threat modeling the high-speed networks that move data between accelerators in distributed AI training clusters.
Read moreEvaluating whether a query-only attacker can poison long-term LLM agent memory and steer behavior across later sessions.
Read moreA small experimental project testing whether misaligned behavior spreads through agent-to-agent interaction in repeated social-dilemma games.
Read moreRecent writing
Threat modeling the high-bandwidth communication layer connecting accelerators, hosts, and storage in distributed AI training.
Read research agendaA small pilot separating context-bound behavioral contagion from persistent trait contagion in repeated agent interactions.
Read findingsWhat query-only memory injection attacks reveal about the security boundary between an aligned model and its persistent memory.
Read research noteA security analysis of fabricated online consensus and the risks it creates for AI systems that synthesize answers from public information.
Read essay