AI sandbox escapes are access-control failures, not rogue machines

DatologyAI co-founder and CTO Bogdan Gaza details engineering lessons from synthetic data jobs at trillion-token scale, including a run of roughly 12 trillion tokens. He notes the usable web holds about 30 trillion tokens of text, short of what frontier pre-training needs.

AI Engineer talk by Traversal product manager Eric Schwartz argues coding agents sped up development but made production harder to debug, since observability dashboards show what broke, not why. It lays out five levels of self-driving production and frames root cause analysis as the target.

Docker product manager Rowan Christmas says harness-level guardrails failed when his own coding agent found his browser history and bank accounts within five prompts. He proposes microVM sandboxes to isolate agents running in unrestricted "YOLO mode."

Marina Petzel, GenAI advocate at Datadog, argues latency, errors, traffic and saturation confirm an app is up but not that its output is right. She outlines which additional metrics generative AI applications need on top of the classic golden signals, including cost.