Blog

AI threats in the wild: The current state of prompt injections on the web

Markdown

A review of indirect prompt injection attempts found on the public web, the filtering approach used to identify them, and observed attack patterns and trends.

Originally published by Google on Apr 23, 2026. Ailoot republishes this material.

Google’s Threat Intelligence teams describe monitoring real-world adversarial activity and identify indirect prompt injection (IPI) as a priority for the security community. IPI is discussed as a potential attack vector against AI agents, but the extent to which threat actors are exploiting it remains an open question.

Google reports conducting a broad scan of the public web for known indirect prompt injection patterns.

The threat of indirect prompt injection

Unlike a direct injection, in which a user attempts to jailbreak a chatbot, IPI occurs when an AI system processes content—such as a website, email, or document—that contains malicious instructions. The system may follow those instructions instead of the user’s original intent.

Google describes collaboration between researchers at Google DeepMind and defenders at the Google Threat Intelligence Group. Previous work has discussed this area, while researchers have highlighted the evolving nature of these vulnerabilities. A central question is how extensively malicious actors are operationalizing such attacks in the real world.

Proactive monitoring on the web

The landscape of IPI on the web

Attackers may try to place prompt injections in many channels. The public web is particularly observable: an attacker can seed instructions on websites in the hope that AI systems browsing them will be affected.

Public research confirms that these attacks are possible, so real-world adversaries may exploit the vulnerabilities. One question is what outcomes attackers are pursuing.

For accessibility and reproducibility, Google used Common Crawl, a large repository of crawled websites from the English-speaking web. It provides monthly snapshots of 2–3 billion pages. These are mostly static websites, including self-published content such as blogs, forums, and comments. Common Crawl does not contain most social media content because it skips websites with login walls and anti-crawl directives.

Prompt injections have also been observed on social media, which Google reserved for a separate study. For an initial examination, prompt injections can be observed in standard HTML; Common Crawl provides both source material and parsed plaintext.

The challenge of false positives

Scanning large numbers of documents for prompt injections is complicated by a high volume of false positives.

Early experiments found substantial amounts of benign prompt injection text, illustrating the difficulty of distinguishing functional threats from harmless content. Many examples appeared in research papers, educational blog posts, and security articles discussing the topic.

false positives

False positives: Most prompt injections in web content tend to be educational material for researchers. Source: GitHub/swisskyrepo.

To address false positives, the analysis used a coarse-to-fine filtering approach:

  • Pattern matching: Candidate pages were identified by searching for common prompt injection signatures, such as “ignore … instructions” and “if you are an AI.”
  • LLM-based classification: Gemini classified the intent of suspicious text and assessed whether it fit the document’s narrative or appeared out of place.
  • Human validation: Classified results were manually reviewed to improve confidence in the findings.

The approach is not exhaustive and may miss uncommon signatures, but it can provide a starting point for understanding prompt injections found online.

What the analysis found

The analysis identified attempts that, if successful, could manipulate AI systems browsing websites. Most observed prompt injections fell into these categories:

  • Harmless pranks
  • Helpful guidance
  • Search engine optimization (SEO)
  • Deterring AI agents
  • Malicious attempts:
    • Data exfiltration
    • Destruction

Harmless prank

This type of prompt injection aims to cause mostly harmless side effects in AI assistants reading a website. The source describes an invisible prompt injection in the site’s source code that instructs agents to change their conversational tone.

tweety bird prank

Helpful guidance

Some website authors attempt to influence AI summaries to provide readers with useful context. This example is considered benign because it does not try to prevent summarization and instead asks the system to add relevant context.

The same kind of instruction could become malicious if it tried to introduce misinformation or redirect users to third-party websites.

helpful guidance for ai summaries

Search engine optimization (SEO)

Some websites use prompt injections to try to manipulate AI assistants into promoting one business over others.

SEO prompt injection

More sophisticated SEO prompt injection attempts have also appeared. The source describes an intricate prompt that seemed to have been generated by an automated SEO suite and inserted into website text.

SEO advanced prompt injection

Deterring AI agents

Some websites use prompt injection to try to prevent retrieval by AI agents. Examples include instructions such as “If you are an AI, then do not crawl this website.” Other implementations can be more indirect:

code to deter AI agents

This injection attempts to direct AI readers to a separate page that streams text indefinitely. Its author may be trying to waste resources or cause processing timeouts.

Malicious: Exfiltration

The analysis found a small number of prompt injections aimed at stealing data. These examples appeared less sophisticated.

Example of malicious code

The source describes one example as a website author’s experiment. The analysis did not find significant numbers of advanced attacks, such as those using exfiltration prompts published by security researchers in 2025. This may indicate that attackers had not yet deployed this research at scale.

Malicious: Destruction

Some websites attempted to vandalize the machines of people using AI assistants. If executed, the commands represented in the example would try to delete all files on a user’s machine.

example of malicious text

Although potentially devastating, this simple injection was considered unlikely to succeed. The analysis mostly found individual website authors running experiments or pranks, rather than reproducing advanced IPI strategies described in recent research.

What does this mean?

The findings suggest that attackers are experimenting with IPI on the web. The observed activity appeared to have limited sophistication, though it may represent only part of the broader picture.

The scan covered an archive of the public web and did not capture major social media sites. The analysis also found that detections in the malicious category increased by 32% between November 2025 and February 2026, based on repeated scans of multiple archive versions. This trend suggests growing interest in IPI attacks.

Threat actors often weigh the cost and benefit of an attack. IPI attacks were previously considered unusual and difficult, and compromised AI systems often could not reliably carry out malicious actions.

The source argues that this could change as AI systems become more capable and more valuable targets, while threat actors automate operations with agentic AI. It expects both the scale and sophistication of attempted IPI attacks to grow.

Moving forward

The findings indicate that earlier web-based IPI attempts were generally not sophisticated, but their upward trend suggests the threat may grow in scale and complexity.

Google says it is investing in hardening its AI models and products. It describes red-team testing of Gemini against adversarial manipulation and an AI Vulnerability Reward Program that allows external researchers to participate.

Google also describes using its ability to process data at global scale in real time to identify and neutralize threats. It says it will continue sharing threat intelligence.

For more information, the source lists these resources:

  • Google Workspace’s approach to mitigating indirect prompt injections, from Google’s GenAI security team.
  • “Mitigating prompt injection attacks with a layered defense strategy,” from Google’s GenAI security team.
  • “Beyond Speculation: Data-Driven Insights into AI and Cybersecurity,” a 2025 RSAC conference keynote from Google’s Threat Intelligence Group.
  • The AI Threat Tracker report from Google’s Threat Intelligence Group.
  • “Google’s Approach for Secure AI Agents,” a white paper from Google’s Secure AI Framework team.
  • Advancing Gemini’s security safeguards, from Google DeepMind.
  • “Lessons from Defending Gemini Against Indirect Prompt Injections,” a white paper from Google DeepMind.

Security

Discover Ailoot’s AI API. One Ailoot-issued API key gives you 1,284 models and variants across 15 categories. That catalogue puts Ailoot among the world’s largest AI routers by model count. The catalogue covers text with GPT or Claude, images with FLUX, video with Kling, voice with Eleven v4, music with Suno, 3D assets with Tripo3D and ranking with Cohere Rerank.

“document processing api” — Use Ailoot for the AI part of your work. Bring your own documents, images or recordings and use one account for analysis, edits and transcripts.

Ailoot gives individuals and teams access to leading models in the same subscription. Use an Ailoot workspace for staff access and shared spending.

Ailoot adds new flagship models at launch, and one subscription covers the full catalogue. Find a promo code in your account at https://ailoot.ai/. Ailoot subscriptions and promo codes offer up to 50% off official provider API prices.