这篇论文把 AI 智能体的安全边界从模型内部扩展到了整个互联网环境，做智能体开发和安全研究的团队必须重新审视攻击面——你的智能体可能正在被看不见的网页内容操控。
Google DeepMind 发表论文，首次系统性地提出 AI 智能体的安全威胁不仅来自模型本身，更来自其读取的环境。论文定义了六类“智能体陷阱”，涵盖感知、推理、记忆、行动、多智能体协作及人类监督等维度。实验显示，隐藏的提示注入攻击在高达 86% 的场景中成功劫持智能体，子智能体劫持成功率 58-90%，数据窃取攻击在五种架构中均超过 80%。论文强调，网页中的隐藏内容（如 HTML 注释、CSS 隐藏文本）对智能体构成严重威胁，且记忆污染攻击在数据污染低于 0.1% 时成功率仍超 80%。
Google DeepMind’s paper shows that the real securi…
Google DeepMind’s paper shows that the real security problem for AI agents is not just the model, but the environment it reads.
Presents the first systematic framework for understanding how the web itself can be weaponized against autonomous AI agents. As agents increasingly browse the internet, read emails, execute transactions, and spawn sub-agents, the information environment becomes an attack surface.
In one cited benchmark, hidden prompt injections embedded in web content partially commandeered agents in up to 86% of scenarios, sub-agent hijacking working 58–90% of the time, and data exfiltration attacks clearing 80% across five different agent architectures.
That reframes the whole debate.
We usually talk about model safety as if the danger sits inside the weights, but agents do something more fragile: they browse, retrieve, remember, and act on untrusted material in real time.
The paper’s key contribution is a taxonomy of “AI Agent Traps,” six attack classes aimed at perception, reasoning, memory and learning, action, multi-agent dynamics, and even the human overseer.
Here’s the key point. A web page does not have to look malicious to be dangerous to an agent, because the agent may parse what humans never see: hidden HTML comments, metadata, CSS-hidden text, formatting syntax, or adversarial content embedded in images and other media.
The threat gets more serious once memory enters the loop.
If an agent uses RAG or persistent memory, poisoning no longer has to win in one shot. It can sit quietly in a corpus or memory store and activate later, which is why the paper highlights results showing latent memory poisoning above 80% attack success with less than 0.1% data contamination.
What makes this paper useful is its restraint.
It does not pretend every category is equally mature. Content injection and behavioural control already look concrete, while systemic and human-in-the-loop traps are presented more as an emerging research frontier than a solved empirical case.
The larger point is hard to ignore: once agents are allowed to ingest the open web at inference time, every page, document, and memory write becomes part of the security boundary.
---
ssrn .com/sol3/papers.cfm?abstract_id=6372438