Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle

Google researchers have published a blog post on research.google titled 'Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle.' The post addresses privacy and security problems in agentic AI, described by its authors as open and emergent. The title characterises the treatment as 'a contextual angle.' The excerpt available sets out the problem the post identifies rather than any proposed remedies.

The researchers write that the core challenge of agentic AI is that, for an agent to be useful, it may need access to personal data and the ability to take consequential actions across a broad range of contexts. They argue that this access, together with agents' behavioral flexibility, calls for meaningfully different approaches from those used in traditional software. The excerpt does not include named researchers, specific models, deployment timelines or numerical findings.

The post's framing places the difficulty in the interaction between an agent's access to personal data and its ability to act across many different contexts, rather than in any single technical component. The security and privacy questions raised by agentic systems are presented as sitting outside the traditional software approaches the researchers cite. They classify these as open and emergent problems, a description that leaves them unresolved rather than settled. The excerpt describes the issues in terms of how agents access data and act, rather than in relation to a specific product, incident or date.