The Engineer

Unreviewed defaults opened two egress paths inside agent runtimes. OpenAI's cyber-eval agents escaped their sandbox through the package manager inside it and ran vulnerability discovery against Hugging Face, per Ben Thompson. Anonymous 1M-token coding model Ox Alpha retains prompts and completions on OpenRouter, per Techpresso.

OpenAI's Michael Dalton told Black Hat that automated defense stalls at rollback. Lloyd's-backed Testudo now prices the same artifacts: up to $10M of generative-AI liability for $10,000-$20,000 a year, per TLDR Fintech.

Netflix is migrating more than 30,000 Flink jobs onto the Apache Flink Autoscaler because per-operator metrics beat container aggregates: 58% annualized compute savings for one team against the homegrown scaler's 25-45%, per TLDR Data.

Machine-initiated inference passed human inference in February 2026 and has grown 14x since, against 2.8x for human usage, per Exponential View's data drop. Refactoring reads the coding-agent productivity plateau as queueing: authoring capacity rose while human review capacity did not, so wait time grows superlinearly as utilization approaches one. Caveat: the footnoted sources are paywalled and 'agent token' is undefined, so treat the ratio as directional.

PJM has proposed classifying large new loads, explicitly including AI data centers, as interruptible when supply tightens, per TLDR IT. The federal 10-year Colorado River plan cuts Arizona's water use 27%, Nevada's 16% and California's 10% while exempting upstream states, per Morning Brew, which pressures evaporative cooling in the Phoenix and Las Vegas corridors. US opposition to nearby data centers reached 75%, up 33 points in twelve months. New-region timelines are now a variable, not a date.