A company uses an AI agent to handle routine administrative tasks like email sorting, meeting transcription, and invoice processing, with a human approval gate for all outgoing communications. The system runs on scheduled jobs with strict rules: alerts only for true urgencies, no autonomous sending, and full audit trails of all operations.
Loopjacking is a security vulnerability where humans approve one operation but agents execute a different one through representation misrepresentation or post-approval state substitution. Researchers demonstrated the attack across multiple agent platforms including Agno AgentOS and LangGraph, while OpenAI's SDK resisted it through serialized continuity checks.