Agents

The risk between expected and unexpected inputs

Route exposed two different meanings for the word "permission"

Real goods, real commitments and no approval of each deal

The OpenClaw flaw and an Ollama scan exposed the limits of “local”

OpenAI's update showed why browser agents need limits beyond just the model

A new control plane, an Australian testing institute and a supplier breach all made assurance more concrete

Anthropic's espionage report showed why agent security needs to understand the full context

What happens when findings come faster than the decision to fix them?

Replit's response showed why instructions need an enforceable boundary around them

What must become stronger as an AI system moves from advice into action