Security of AI Agents

Key findings

• Identifies session management, model pollution, privacy leakage, agent programs and sandboxing as core vulnerability areas - Proposes defences at each layer

Sources

• arxiv.orghttps://arxiv.org/abs/2406.08689