Yuki Tanaka-Osei

Attackers exploit every stage of retrieval pipelines to steal sensitive data.

Runtime isolation for AI-generated code requires VM boundaries plus controls above the sandbox.

How to build logging that captures what agents do and why it matters.

Human judgment of agent decisions must become an explicit SLO metric in on-call rotations.

LLM outages split into four distinct failure modes, each requiring different defenses.

Automated safeguards prevent runaway agents from draining budgets undetected.