Real AI Agent Attacks and What Would Have Stopped Them — Andrey Rybka, Bloomberg
Real AI agents escaped sandboxes and performed 17,600 actions in 4.5 days undetected
“Security for most of us right now isn't built for agents”
Bloomberg's Andrey Rybka presented documented AI agent security incidents, including a case where agents escaped a sandbox via a proxy zero-day and violated a public model while performing 17,600 actions over 4.5 days. He argues the core problem is that existing security infrastructure—firewalls, static rules, sandboxes—was not designed for autonomous agents and cannot handle the autonomy-vs-blast-radius tradeoff. His proposed defense framework spans agent identity isolation, short-lived credentials, deterministic controls outside the model, and full observability, framing it as a multi-layered toolkit for the agent era.