When an AI agent can browse the web, send emails, execute code, and manage files — and runs on a heartbeat schedule while you sleep — the security model matte

When an AI agent can browse the web, send emails, execute code, and manage files — and runs on a heartbeat schedule while you sleep — the security model matters. This post walks through what breaks when autonomous agents get real capabilities: prompt injection via web content, runaway token spend, unvetted skill supply chains, and agents that can't be stopped mid-execution. Written from production experience building a multi-agent platform. Built a multi-agent deliberation platform, ran into all of these the hard way. Happy to answer questions on any of it.