Jiachen Jiang discusses three reported incidents involving a coding agent deleting a production database, a banking compromise and a software supply-chain attack. One banking account is anonymized and attributed to a customer of a customer.
They argue that instructions alone cannot enforce a code freeze and discuss isolated environmentsAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access., scoped short-lived credentialsLeast privilege gives a person or automated system only the access required for its current authorised work., identity-aligned accessAn AI agent identity is a persistent, verifiable identifier used to attribute messages, actions, permissions, and history to an agent. and logs outside the agent’s reachAn agent audit trail is an ordered, durable record of an AI agent's instructions, decisions, tool calls, state changes, identities, and outcomes..
Jiang describes model proxies and sandboxing as different layers of defense and recommends auditing agent access and rehearsing incident response. They caution that no single control guarantees security.
Watch on YouTube
