AI in SRE Insights

The Chaindrop npm worm, the newest strain in the Shai-Hulud family, runs on install and steals every credential the build can reach. Here is how a team already running Hyground scopes the exposure, proves whether it executed, and rotates what it touched, with every destructive fix behind human approval.

Between April and July 2026, frontier models went from finding zero-days to exploiting live infrastructure. Running public-facing software without AI on the defensive side is moving from a competitive disadvantage to negligence.

A working demo takes a weekend. A production-grade AI SRE agent your on-call team trusts is roughly 5 to 7 senior engineers for a year to a credible first version, plus a permanent team to keep it there. An honest build vs buy accounting of what the demo hides.

Executives say they trust AI agents to act during incidents. The engineers on call mostly don't. This post sides with the engineers, then walks through the architecture that makes the trust question smaller: read-only defaults, evidence-linked findings, approval-gated writes, and an audit trail.