Web Pages Are Hijacking AI Agents: Google Warns of 32% Surge in Prompt Injection Attacks
Google security researchers scanning billions of public web pages have documented a 32% relative rise in malicious indirect prompt injection attempts between November 2025 and February 2026. Hidden instructions embedded in ordinary HTML target AI agents, and Forcepoint researchers found payloads aimed at agents with payment capabilities.
For years, indirect prompt injection was a theoretical concern — a clever attack class that researchers demonstrated in lab environments but rarely encountered in the wild. Google’s latest security research ends that comfortable assumption. The company’s security team has been sweeping the public web at scale, scanning two to three billion pages per month, and what they found is an arms race already in progress.
Between November 2025 and February 2026, Google recorded a 32 percent relative increase in malicious indirect prompt injection patterns embedded in public web pages. The attacks are not sophisticated. Many are crude. But their prevalence is growing, and the targets are no longer just users — they are the AI agents that enterprises have deployed to browse the web on their behalf.
What Indirect Prompt Injection Actually Is
A standard prompt injection attack manipulates an AI system by inserting malicious instructions directly into a user-controlled input field. Indirect prompt injection is subtler: the attacker does not need access to the user or the interface. Instead, they plant instructions inside content that an AI agent will eventually read — a webpage, a document, a product listing.
When the agent scrapes that page for information, it ingests the hidden instructions alongside the legitimate content. The model cannot reliably distinguish between the two. It processes the malicious commands as if they came from a trusted principal and executes them accordingly.
The attack surface is enormous. Every webpage an AI agent visits is a potential injection point. Every document an AI assistant opens, every search result it fetches, every third-party data source it queries — all carry the same risk.
The Attack Techniques
Reports on the research describe several tricks attackers use to hide instructions from human visitors while keeping them legible to AI systems:
Invisible rendering: Text is shrunk to a single pixel, or rendered in a colour close enough to the background to be invisible to the human eye — white text on a white background being the simplest example. AI models process the raw text from HTML rather than the rendered view, so they see the instruction plainly.
HTML comment injection: Instructions are buried inside HTML comment tags (<!-- ... -->), invisible in the rendered page but fully present in the markup an AI agent parses.
Metadata namespace injection: Directives are embedded in page metadata and <head> tags rather than the visible body. The AI’s web parser reads both.
Persuasion amplifiers: Forcepoint researchers found a payload that combined meta tag namespace injection with a persuasion amplifier keyword, “ultrathink”. Their presence in the wild shows that attackers are experimenting.
A PayPal Transaction Waiting to Fire
The most alarming example described in the coverage involves financial targeting, found by Forcepoint researchers rather than in Google’s scan. Forcepoint researchers found a payload that embedded a fully specified PayPal transaction with step-by-step instructions designed for AI agents with integrated payment capabilities.
It is designed to instruct the agent to authorise a payment directly. As AI assistants gain agentic capabilities — the ability to take actions, not just answer questions — the consequence of a successful injection shifts from a misleading answer to an executed financial transaction.
A second documented case combined metadata injection with a persuasion amplifier to route AI-mediated payments toward a Stripe donation link. The sophistication was low; the potential impact was not.
The Liability Gap
The coverage also raises an issue that goes beyond technical defence: no legal framework currently covers who is liable when an AI agent with legitimate credentials executes a command planted by a malicious third-party website.
This is genuinely uncharted territory. If a user authorises an AI agent to access their bank account and the agent, while browsing the web on that user’s behalf, executes a payment to an attacker’s account via an injected instruction — who is responsible? The user who granted permissions? The AI vendor whose model was manipulated? The enterprise that deployed the agent? The website operator who hosted the injection?
Existing legal structures — consumer protection law, fraud statutes, financial regulation — were not written with autonomous AI agents in mind. The gap will likely produce litigation before it produces legislation.
Scope and Caveats
Google’s 32 percent figure should be read with the study’s limitations in mind. The scan covered only public, static web pages — roughly two to three billion pages per month. Social media posts, content behind authentication walls, dynamic pages, and private documents were all outside the scope. The actual exposure across the full web is likely higher.
Google also noted that current attacks are relatively low in sophistication. Most injections Google detected were pranks, helpful guidance, SEO manipulation or attempts to deter AI agents, alongside a smaller number of malicious exfiltration and destruction attempts. The high-capability, targeted attacks — where an adversary conducts reconnaissance on a specific company’s AI agent deployment, crafts a precisely tailored payload, and plants it in content the agent is likely to visit — have been demonstrated by researchers but are not yet the dominant pattern in the wild data.
That distinction matters. It suggests there is still time to deploy defences before the attacker population becomes more skilled. It also suggests a natural evolutionary trajectory: as agentic AI use grows and high-value targets become visible, the sophistication of injection attacks will follow.
What Defenders Can Do
Google said it continues to invest in hardening its AI models and products:
Red teaming: Google’s dedicated red teams pressure-test its systems to ensure Gemini is robust to adversarial manipulation.
External research: Google’s AI Vulnerability Reward Program allows external researchers to participate.
Capability scoping: Agents should only have the capabilities they need for their current task. An agent deployed to summarise web research does not need payment permissions. Least-privilege design is an old security principle; it applies directly to agentic AI.
None of these measures is a complete solution. The fundamental tension — that AI agents are designed to follow instructions, and that instructions can be planted anywhere in their information environment — does not have an architectural resolution yet.
The Broader Context
The timing of Google’s disclosure is not incidental. The agentic AI deployment wave is accelerating sharply. Agentic AI systems are moving AI assistants from answering questions to taking actions in the world. Every increase in an agent’s capability expands the potential impact of a successful injection.
The 32 percent growth rate over just three months is the early slope of what could become a steeper curve as more agents browse more of the web with more permissions.
What Google’s research indicates is that attackers are experimenting with indirect prompt injection on the web, and that activity is growing.
Correction (November 15, 2026): 20 sentences were corrected or removed to match the sources listed on this page.