How we monitor internal coding agents for misalignment | AI Sec Watch