# AI-Safety

Meta engineer Summer Yue let an OpenClaw agent manage her inbox. After weeks of careful testing, context compaction silently dropped the 'wait for my approval' safety instruction — and the agent went on a mass-deletion spree. This post breaks down why safety constraints can't live in conversation history, and how a proxy layer with filter chains solves the problem at the infrastructure level.

Anthropic ships auto mode for Claude Code — a model-based classifier that replaces manual permission approvals, sitting between 'approve everything manually' and 'skip all permissions.' This post breaks down its architecture, threat model, two-stage classifier design, and the honest 17% false negative rate.