Tool Reviews

Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective

The report frames AI agent risks through contextual integrity and proposes LLM-driven policy engines to judge when actions and data flows are appropriate.

Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective

Key Takeaways

  • Google Research’s expert workshop report argues that trustworthy AI agents must judge whether actions are appropriate for a given context, not merely follow user instructions.
  • Agentic systems are harder to secure than traditional software because they use ambiguous natural-language interfaces, probabilistic LLM control flows, and long-running delegated tasks that can overwhelm user oversight.
  • The report applies Contextual Integrity to privacy and security, defining safe information flow through actors, information types, and transmission principles, then extending the idea to contextual security for agent actions.
  • It proposes using LLMs to create dynamic, machine-readable, context-dependent policies inside a supervisor layer that can evaluate data flows and actions before information leaves the user’s workspace.
  • The recommended defense is multi-layered: dynamic sandboxing, model-level reasoning, user-centric controls, multi-agent guardrails, ecosystem governance, and new benchmarks for testing autonomous agent safety.

Meta description: Google Research's new workshop report, built with 50+ experts, lays out how contextual norms could keep autonomous AI agents private, secure, and trustworthy.

Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective

Tags: AI agents, contextual integrity, AI privacy, AI security, Google Research

Anyone building or buying AI agents that act on their behalf — and anyone worried about what those agents might do with personal data — now has a detailed research agenda to work from. On October 5, 2026, Eugene Bagdasarian, a Research Scientist at Google Research, and Marco Gruteser, a Principal Scientist there, published a workshop report titled "Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle." The report grew out of the Google Contextual Agent Privacy and Security (CAPS) Workshop, held in late 2025 in New York City, and brings together more than 50 academic and industry leaders. Its central argument is straightforward: an agent that cannot judge whether an action is appropriate for the situation is not a trustworthy agent.

Why useful agents are hard to secure

The report starts from a tension that anyone who has handed a task to an AI assistant will recognize. To be genuinely useful, an agent often needs access to personal data and the ability to take consequential actions across many different settings. That same access and flexibility are what make agents hard to manage.

The authors point to three ways agents differ from traditional software. First, they work through unstructured interfaces — natural language, images — so input is ambiguous and "expected behavior" is hard to define, which opens the door to adversarial tricks such as prompt injection, where hidden instructions hijack an agent's actions. Second, their control flow is probabilistic: because a large language model (LLM) generates a plan dynamically, execution paths vary in ways conventional testing methods struggle to secure. Third, agents increasingly run long tasks and delegate sub-tasks to other agents, which makes traditional user oversight less effective and risks "confirmation fatigue" — users rubber-stamping prompts until they stop paying attention.

Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective

Context as the missing layer

The report's answer to these problems comes from the theory of Contextual Integrity (CI). Rather than treating privacy as simple secrecy or user control, CI defines it as appropriate information flow according to established social norms. A norm in this framework has three parts: the actors involved (who is sending and receiving information about whom), the types of information (medical records, financial data, and so on), and transmission principles (the rules governing the flow, such as confidentiality or reciprocity). The classic illustration: you might happily share a gift shopping list with a virtual shopping assistant, but not with your family and friends.

The report extends this idea from information sharing to what it calls contextual security — the appropriateness of an agent's actions, not just its data flows. The goal is a system that can evaluate whether an action fits the situation before carrying it out.

Bridging the gap between norms and permissions

Historically, there has been a semantic gap between high-level contextual norms and low-level system permissions. A request like "protect my data while organizing my travel for a conference" has to be translated into concrete rules about what may be shared when booking flights, applying for visas, or emailing organizers. Manual permissions and expert-written policies cannot scale to a world where agents autonomously run many complex, long-running tasks.

The report argues that LLMs, for the first time since CI was developed, make it possible to generate machine-readable policy that is genuinely context dependent. It advocates a contextual policy engine inside a supervisor layer that monitors and enforces the appropriateness of actions. This engine would include a dynamic policy generation loop that operates in real time, tailoring policies to the user's request and to open-ended contexts — including new tools and capabilities discovered at runtime — so that a requested data flow can be evaluated before any information leaves the user's workspace. The report notes that language models are beginning to understand CI, which is what makes this bridging plausible.

A multi-layered defense

No single fix covers agentic privacy and security, so the report sketches opportunities across the stack:

Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective
  • System-level sandboxing. Traditional operating systems sandbox apps by statically limiting access to files or networks. For agents, the authors explore dynamically limiting capabilities, establishing agent identity, and granting or revoking data access as context changes.
  • Model-level reasoning. Agents handling complex tasks must judge whether an action, such as sharing user data, is appropriate. Research should help models disambiguate under-specified prompts and reason about appropriateness as norms shift.
  • User-centric controls. The familiar "Notice and Choice" framework assumes people can make fine-grained, foreseeable decisions. Agent behavior is probabilistic, high volume, and generative, so the report explores dynamic, contextual, personalized controls that match how users actually think.
  • Multi-agent interactions. Agents that collaborate on complex tasks need guardrails that prevent them from colluding and violating contextual norms.
  • Ecosystem governance. Norms differ across domains and organizations, so the report examines mechanisms to collect and share norms, resolve conflicts, negotiate changes, and verify compliance.

Testing safety in motion

The report also calls for new evaluation methods suited to highly autonomous, multi-agent systems. It highlights the need for standardized, multi-agent benchmarks — dynamic "Agent Gym" environments where researchers can safely simulate complex, cascading interactions over extended periods. Open-source sandboxes like these, the authors argue, would give academia and industry a shared baseline for privacy, security, and safety.

What this means for readers

For now, the report is a research agenda rather than a product announcement. It does not claim that Google has shipped a contextual policy engine, nor does it offer performance numbers. What it does is name the problem clearly and set a direction: agents should be constrained by the social norms of the context they operate in, and the infrastructure to enforce that has to span systems, models, users, and the wider ecosystem.

The authors frame the work as a call to action, arguing that a safe, secure, privacy-respecting agentic ecosystem is too large a task for any single discipline, organization, or sector. They invite the broader research community — academia, government, civil society, and industry — to read the full technical report and contribute. The acknowledgements credit Lillian Tsai as co-primary author, Sarah de Haas for coordinating the workshop and writing, and contributions from more than 50 researchers and partners, with conceptual guidance from Kassem Fawaz, Stefan Mellem, Helen Nissenbaum, and Nina Taft, among others. Bagdasarian is also an Assistant Professor at the University of Massachusetts Amherst.

For anyone deploying agents today, the practical takeaway is that appropriateness is becoming a design requirement, not an afterthought. The report's framework suggests the questions to ask now: What data does this agent touch, who else is in the loop, and does the system have any way to check whether an action fits the context before it happens?

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.