Google Research has released a comprehensive framework addressing privacy and security risks in autonomous AI agents, drawing on findings from a major industry workshop with over 50 experts. The report identifies three critical challenges where AI agents diverge from traditional software: their processing of unstructured inputs like natural language and images (vulnerable to prompt injection), probabilistic execution paths that bypass conventional testing, and increasing autonomy that can overwhelm human oversight through "confirmation fatigue." To address these risks, researchers propose grounding agent security in Contextual Integrity theory, which treats privacy not as absolute secrecy but as appropriate information flow according to social norms. An agent might be permitted to share travel details with a booking assistant while restricting the same data from social contacts, reflecting contextual appropriateness rather than blanket data protection. The team advocates for a contextual policy engine that operates as a supervisor layer, dynamically generating and enforcing policies based on user requests and runtime conditions. This architecture would evaluate whether an action is socially appropriate before execution, potentially solving the long-standing gap between high-level privacy norms and low-level system permissions in complex, autonomous AI systems.