GoKawiilQueen's University Belfast researchers flag zero-trust gaps in AI agent systems
Researchers from Queen's University Belfast describe how multi-agent AI systems, which use orchestrators to manage task forces of agents for jobs like risk assessment and client profiling, maintain logs and checkpoints meant to delete historic data to prevent biased decisions. They warn that despite these deletion checkpoints, future prompts can trigger a 'context recovery attack,' where agents resurface supposedly deleted information. The analysis also raises concerns about zero-trust architecture components, such as the policy enforcement point, and whether their enforced policies can themselves be trusted.