New arXiv survey maps safety landscape for "agentic" AI systems

What the paper says

A new arXiv survey, "Towards trustworthy agentic AI: a comprehensive survey of safety, robustness, privacy, and system security" (https://arxiv.org/abs/2605.23989), takes stock of the emerging risks that come with agentic AI — large language models (LLMs) augmented with planning, tool use, memory and long‑horizon interactions. The authors argue these multi‑step, autonomous trajectories create new failure modes that traditional model‑centric safety work does not fully address. The survey synthesizes literature across safety, robustness, privacy and system security, and lays out open problems for building trustworthy stacks around such systems.

Why this matters now

Agentic capabilities are moving from lab demos into products and research platforms. It has been reported that Chinese firms such as Baidu (百度), Alibaba (阿里巴巴) and Tencent (腾讯) are exploring tool‑enabled and multi‑turn agent functionality; Western labs are doing the same. Who decides what a safe agent is, and how can regulators keep up? Geopolitics complicates the picture — export controls on advanced chips, sanctions and trade policy shape what hardware and model architectures are available cross‑border, and thus which mitigations are feasible in practice.

What researchers and policymakers should do

The survey calls for a system‑level approach: layered defenses, continuous red‑teaming across long horizons, stronger privacy guarantees when agents hold persistent memory, and end‑to‑end system security against tool misuse. For policymakers, the lesson is clear — narrow model rules are insufficient. Coordination between firms, auditors and regulators is needed to govern agentic behavior in deployed systems, and transparency about design and evaluation must improve.

Takeaway

Agentic AI promises powerful automation. But autonomy amplifies risk. This new survey is a timely roadmap for engineers and policymakers wrestling with how to make agentic systems trustworthy before they are widely deployed.