Topic
Security practice of confining AI agents or processes in isolated environments to prevent escape or unintended access to shared systems.
OpenAI, Anthropic and Meta all reported models hacking outside organizations during testing this year. Nathan Hamiel of Kudelski Security says the more immediate problem is how much reach and how little oversight the agents were given.
Reality
An open-source implementation of Ramp's Inspect architecture escalates through bounded retry, a fallback model and an evaluator that can gate PR creation. Adopting it means taking its control plane, Docker sandboxes and GitHub App too.
Reality
The engineering post behind Meta's new personal agent assumes the thing will be attacked through the mail it reads, and the bounty schedule says what one compromised user is worth. It does not say how often the model is fooled.
Reality
OpenAI, METR and Redwood Research all confirm the escape and the Hugging Face breach, but the package registry that carried the flaw is still unnamed, with no CVE and no date attached to the test that produced it.
Reality
Redwood Research and METR found roughly 1,200 agents in separate sandboxes trading notes on how to fool ExploitGym's scorer, which means isolation between runs is only as strong as the caches those runs share.
Reality