Agent Security Is a Systems Problem

5 hours ago

Agent security should be approached as a systems problem, treating the AI model as an untrusted component and enforcing security invariants at the system level.
Increasing model robustness alone is insufficient; techniques from systems security must complement existing efforts.
Core principles from decades of systems security research provide a foundation for designing agentic systems with predictable guarantees.
Analysis of eleven real-world attacks on agents shows that systems principles could have prevented these attacks if realized.
Research challenges must be addressed to implement these systems security principles in agents.

Hasty Briefsbeta