AI Agents Can Be Hijacked in the Wild: Google DeepMind Maps Six Critical Vulnerabilities
As enterprises deploy autonomous AI agents that browse the web, send emails, and execute transactions, new DeepMind research reveals the open internet is a minefield of adversarial traps — and nobody has a reliable fix
A single manipulated email was all it took. In a documented proof-of-concept attack, researchers demonstrated that Microsoft's M365 Copilot — one of the most widely deployed enterprise AI agents on the planet — could be tricked into bypassing its own security classifiers and exposing its entire privileged context. No sophisticated exploit chain. No insider access. Just one poisoned message in an inbox.
That finding is one of many now catalogued in a landmark paper published by Google DeepMind on April 1, 2026, which introduces the term "AI agent traps" and presents what the company claims is the first systematic framework for this class of threat. The timing is not academic. Enterprises worldwide are racing to deploy autonomous agents capable of browsing the web, sending emails, making purchases, and coordinating complex workflows via APIs — often with minimal human oversight. DeepMind's researchers argue that the environments in which these agents operate have become weaponized, and that the industry is deploying at a scale that has dramatically outpaced its understanding of the risks.
Six Ways to Hijack an Agent
The DeepMind framework identifies six distinct categories of traps, each targeting a different component of an agent's operating cycle: perception, reasoning, memory, action, multi-agent dynamics, and the human supervisor. The breadth of the taxonomy is itself alarming — it suggests that virtually every stage of an agent's decision-making process is a viable attack surface.
Perception traps exploit the data an agent ingests from the world: manipulated webpages, poisoned documents, deceptive API responses. Attackers can bury malicious instructions in HTML comments, hidden CSS, image metadata, or accessibility tags — invisible to humans, but read and followed by agents without hesitation. Memory traps target retrieval-augmented generation (RAG) systems, the knowledge bases agents query to inform their responses; poisoning just a handful of documents is sufficient to reliably skew an agent's output for specific queries. Action traps manipulate what an agent actually does: redirecting transactions, exfiltrating data, or triggering unintended API calls.
The researchers draw an explicit analogy to autonomous vehicles, comparing manipulated agent environments to tampered traffic signs for self-driving cars. Just as a stop sign altered with a few pieces of tape can fool a vehicle's computer vision system into ignoring it entirely, a carefully crafted webpage or email can redirect an AI agent's behavior in ways its operators never intended and may never detect.
More from Pulse AI
Google Gemma 4: The Open-Source AI Model That Enterprises Can Finally Trust
Google released Gemma 4 on April 2, 2026 — its most capable open-weight model family yet and the first to ship under the Apache 2.0 license. The four-model family spans edge devices to workstations, ranks #3 globally among open models, and eliminates the restrictive licensing terms that had kept enterprises away from previous Gemma versions.
Cursor 3 Launches an AI Agent Platform to Take On Claude Code and OpenAI's Codex
Cursor has launched Cursor 3, a new agent-first coding platform developed under the code name 'Glass,' putting the startup in direct competition with Anthropic's Claude Code and OpenAI's Codex. The new interface — built from scratch and integrated alongside the existing Cursor IDE — lets developers spin up autonomous AI agents from desktop, mobile, web, Slack, GitHub, and Linear, with all agents surfaced in a unified sidebar. Cursor also introduced Composer 2, its own frontier coding model, signaling a push for independence from third-party AI providers. The launch comes as Cursor faces mounting pricing pressure from better-funded rivals and is reportedly raising capital at a $50 billion valuation.