2 months 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.