Artificial intelligence agents are becoming more capable of browsing websites, analyzing documents, reading emails, accessing business applications, and performing tasks on behalf of users. These capabilities create new opportunities for productivity, but they also introduce a security problem that traditional defenses were not designed to handle.

One of the most concerning risks is indirect prompt injection: an attack in which malicious instructions are embedded in content an AI system processes, potentially causing it to ignore its intended task, disclose sensitive information, or perform unauthorized actions.

In April 2026, the Cloud Security Alliance (CSA) published research examining the growing evidence of indirect prompt injection activity in real-world environments. Google's security research reported a 32% relative increase in detected malicious indirect prompt injection content between November 2025 and February 2026.

These findings do not mean every AI agent is vulnerable or that every detected instruction results in a successful attack. They do, however, demonstrate why organizations need to treat prompt injection as a practical security concern rather than solely a laboratory experiment.

The central question for businesses is no longer just whether an AI model can recognize malicious instructions. It is whether the surrounding system can prevent those instructions from causing harm.

Key Takeaways

  • A growing security concern: Research published in 2026 documents malicious prompt injection activity on the web and cases involving AI agents.

  • Google's reported increase: Google observed a 32% relative rise in malicious indirect prompt injection content between November 2025 and February 2026.

  • AI agents expand the attack surface: Systems that can access private data and perform external actions face greater potential consequences when manipulated.

  • Prompt instructions are not enough: Security depends on permissions, tool restrictions, isolation, monitoring, and approval workflows.

  • Human oversight remains important: Sensitive or irreversible actions should require appropriate authorization.

  • Security reporting has limitations: A vulnerability does not always receive a CVE identifier, so organizations should monitor multiple sources of security intelligence.

What Is Indirect Prompt Injection?

Indirect prompt injection occurs when an AI system encounters malicious instructions inside external content rather than receiving them directly from the user.

That content might appear in a website, email, uploaded document, support ticket, shared file, or other material the system has been asked to process.

For example, imagine that an employee asks an AI assistant to summarize a collection of webpages. One page contains instructions intended to manipulate the assistant into revealing information from its conversation or attempting an unauthorized action.

The employee may never see those instructions. Nevertheless, the AI system might interpret them as relevant directions instead of treating them as untrusted content.

Whether the attack succeeds depends on the model, the application's architecture, its permissions, the tools available to it, and the security controls surrounding its actions.

This distinction matters because an AI agent can create risks beyond an incorrect answer. If it has access to email, cloud storage, databases, or external communication tools, a successful manipulation could potentially affect real systems.

Direct vs. Indirect Prompt Injection

CharacteristicDirect Prompt InjectionIndirect Prompt Injection
Source of malicious instructionsUser-provided inputExternal content processed by the AI
Possible delivery methodChat messages or submitted promptsWebsites, emails, documents, and retrieved content
User awarenessThe user generally sees the submitted inputThe user may not notice the embedded instructions
Main security concernManipulation of the model's intended behaviorManipulation through content the system retrieves or processes
Relevant defensesInput handling, instruction hierarchy, tool restrictionsModel safeguards, content isolation, least privilege, and action controls

Neither type is automatically successful. The distinction helps security teams identify where untrusted instructions enter an AI workflow and which controls are needed to contain them.

What the 2026 Research Reveals

The Cloud Security Alliance's research note, Indirect Prompt Injection Goes Operational, brings together findings from Google, Forcepoint X-Labs, Palo Alto Networks Unit 42, and other security research.

The research describes a developing threat landscape in which malicious instructions are being placed in content that AI systems may encounter during normal operations.

1. Google Reported a 32% Relative Increase

Google reported a 32% relative increase in malicious indirect prompt injection content detected between November 2025 and February 2026.

Its analysis covered versions of the Common Crawl web archive. Google also described a crawling operation that processes approximately 2–3 billion pages per month.

The figure requires careful interpretation.

A 32% relative increase does not mean that 32% of webpages contain malicious instructions. Nor does it mean that successful attacks against AI agents increased by 32%.

Instead, it describes an increase in the malicious category identified by Google's analysis over the specified period.

The findings indicate growing activity, but they do not establish the overall prevalence of successful attacks across every website, AI product, or enterprise environment.

2. Security Researchers Have Documented Real-World Cases

The CSA research note also references investigations into malicious prompt injection activity by security research teams.

Forcepoint X-Labs catalogued verified injection payloads found in the wild, while Palo Alto Networks Unit 42 documented cases involving AI agents and multiple delivery techniques.

These findings help move the discussion beyond hypothetical demonstrations. Researchers are investigating malicious content intended to influence AI systems operating in real environments.

However, detecting a malicious payload, demonstrating that an agent can be manipulated, and confirming a successful compromise are different evidentiary categories. Security reports should distinguish between them.

For organizations deploying AI agents, the practical lesson is to prepare for attempted manipulation even when the success rate of a particular attack remains uncertain.

3. The Risk Increases When AI Systems Can Take Action

An AI assistant that only generates text has a different risk profile from an agent that can access sensitive files, send messages, update records, or interact with external services.

Consider a business assistant with permission to read customer emails and draft replies.

A malicious instruction embedded in an email could attempt to redirect the assistant's behavior. If the application also gives the agent unrestricted access to customer records and permission to send messages automatically, a successful manipulation could have consequences beyond the original email.

The risk depends on the system's actual capabilities and safeguards. Reading malicious content does not automatically grant an attacker access to the user's account.

Nevertheless, the more authority an agent possesses, the more important it becomes to limit what that agent can do.

Why AI Agents Create a Different Security Challenge

Traditional applications generally execute operations according to predefined code paths and explicit application logic.

Large language models, by contrast, process natural-language instructions alongside information retrieved from external sources. They can be influenced by text that appears relevant to their current task, including text that was never intended to be trusted.

This creates a challenge for systems that combine language models with external tools.

An agent might be asked to research a topic, read several documents, compare findings, and then send a report. During that process, it may encounter content containing instructions that conflict with the user's original request.

If the application does not properly constrain the agent, the malicious content may influence its decisions.

Three conditions are particularly important:

  • Access to sensitive information: The agent can retrieve information that should remain confidential.

  • Exposure to untrusted content: The agent processes material that an attacker may be able to influence.

  • Ability to communicate or take action: The agent can send data elsewhere or modify an external system.

Security researcher Simon Willison describes this combination as the lethal trifecta of AI agent security.

The framework explains why prompt injection can become especially dangerous when data access, untrusted inputs, and external communication are combined.

When these three conditions overlap, an AI agent may have the capabilities needed for an indirect prompt injection attack to cause harm. The diagram below illustrates how private-data access, exposure to untrusted content, and outward communication can combine to create a security risk.

Diagram of the lethal trifecta of AI agent risk: access to private data, exposure to untrusted content, and outward communication combine so an agent can be steered by injected instructions. Mitigations shown: privilege separation, human confirmation for irreversible actions, treating retrieved content as data, and logging agent actions

An application does not need to eliminate every possible malicious instruction to reduce this risk. It needs to prevent untrusted content from gaining the authority required to access sensitive information or execute harmful actions.

Where Can Indirect Prompt Injection Appear?

Indirect prompt injection is not limited to a particular website type or AI product. It can arise wherever an AI system processes content that an attacker can influence.

Potential entry pointExample riskRecommended control
Public webpagesRetrieved content attempts to redirect an AI research taskTreat page content as untrusted data
EmailsA malicious message attempts to influence an email assistantSeparate reading permissions from sending permissions
Uploaded documentsEmbedded instructions attempt to change document-processing behaviorIsolate document content and restrict available tools
Customer support ticketsUser-submitted text attempts to manipulate an automated support agentApply strict authorization and action policies
Shared filesCollaboratively edited content contains malicious instructionsEnforce access controls and validate file sources
Retrieval-augmented generation systemsRetrieved passages contain instructions that conflict with the user's requestApply retrieval controls, content boundaries, and output validation
Coding assistantsRepository content attempts to influence code changes or tool useUse sandboxing, limited permissions, and human review

These examples describe potential attack paths, not proof that every system in each category is vulnerable.

The important question is whether an attacker can place content where an AI agent will encounter it and whether that content can influence actions the agent is authorized to perform.

The Most Important AI Agent Security Defenses

There is no single control that reliably eliminates indirect prompt injection. Effective protection requires multiple layers that limit both the likelihood of manipulation and the damage a successful attack could cause.

The following recommendations align with established application-security principles and the OWASP GenAI Security Project's treatment of prompt injection.

1. Apply the Principle of Least Privilege

An AI agent should receive only the permissions necessary to complete its assigned task.

An assistant that summarizes webpages does not normally need permission to send emails, delete files, or transfer money.

Likewise, an agent that drafts a report should not automatically receive unrestricted access to every internal database.

Separate permissions by task, limit access to sensitive information, and avoid granting broad administrative privileges to general-purpose agents.

Why it matters: Even if malicious content influences the model, the agent's limited authority can prevent that manipulation from becoming a serious incident.

2. Separate Untrusted Content From Authorized Instructions

Retrieved webpages, emails, documents, and other external material should be treated as data rather than trusted instructions.

Applications should clearly distinguish the user's task and the system's security rules from the content being analyzed.

Structured content boundaries, trusted instruction hierarchies, and explicit policies can help reduce confusion. However, formatting and prompting alone cannot guarantee that a model will always ignore malicious instructions.

The application should enforce critical restrictions outside the model whenever possible.

Why it matters: The system should not depend entirely on the model correctly identifying every malicious instruction.

3. Require Approval for Sensitive Actions

Actions such as sending external messages, publishing content, deleting important files, changing permissions, or initiating payments may require explicit authorization.

Design approval workflows according to the potential impact of each action.

For example, an AI agent could prepare an email and recommend a recipient without having permission to send it automatically. A user could then inspect the message and approve it.

Low-risk actions may be automated, while sensitive or irreversible operations require stronger checks.

Why it matters: Human approval creates an opportunity to detect suspicious behavior before the action is completed.

4. Validate Tool Calls Outside the Model

An AI model should not be the sole authority deciding whether an operation is permitted.

The application should independently verify the requested action, the user's authorization, the target resource, and any relevant business rules.

For example, an agent requesting access to a customer record should be subject to the same access-control requirements that apply to other application components.

Tool permissions should be enforced by the system executing the action, not merely described in the model's instructions.

Why it matters: Independent enforcement can prevent a manipulated agent from exceeding its authorized capabilities.

5. Isolate High-Risk Workflows

Use sandboxes and separate execution environments for tasks that involve untrusted documents, code execution, or other potentially dangerous operations.

Limit access to the host filesystem, network connections, credentials, and unrelated applications.

Where possible, separate research and analysis from the systems that execute consequential actions.

Why it matters: Isolation reduces the potential impact of a failure in one component.

6. Log and Monitor Agent Activity

Security teams should record relevant agent operations, including tool calls, authorization decisions, access to sensitive resources, and consequential actions.

Logs should help investigators reconstruct what happened without unnecessarily storing passwords, confidential content, or other sensitive information.

Monitoring can also identify unusual behavior, such as repeated attempts to access unrelated resources or unexpected outbound communications.

Why it matters: Without reliable records, organizations may struggle to detect, investigate, or recover from an incident.

7. Test the Complete Application, Not Just the Model

A model performing well on a prompt-injection benchmark does not prove that an application is secure.

Testing should examine the entire workflow, including retrieval systems, tool integrations, permissions, approval mechanisms, and output handling.

Security teams should test whether malicious content can influence actions and whether the application blocks unauthorized operations even when the model behaves unexpectedly.

Testing should be repeated when the model, tools, permissions, or workflow architecture changes.

Why it matters: Real-world security depends on the entire system rather than on model behavior alone.

8. Track Security Advisories Beyond CVE Feeds

A CVE identifier is useful for tracking a disclosed vulnerability, but not every security incident or AI-related weakness receives one.

Organizations should also monitor vendor advisories, security research, responsible disclosures, and relevant OWASP publications.

Maintain an inventory of AI systems, their connected tools, the data they can access, and the actions they can execute. Use that inventory to prioritize security reviews and remediation.

Why it matters: A CVE-only monitoring strategy can miss important developments involving application design, unsafe permissions, and prompt injection.

A Practical Checklist for Organizations Using AI Agents

Before deploying an agent that interacts with external content or business systems, security and engineering teams should be able to answer the following questions.

  • Does the agent have only the permissions required for its task?

  • Can it access confidential data that is unrelated to the task?

  • Can untrusted content influence the tools the agent chooses to use?

  • Are authorization rules enforced independently of the language model?

  • Do sensitive or irreversible actions require appropriate approval?

  • Are file access, network access, and code execution restricted?

  • Are important tool calls and security decisions logged?

  • Can the system detect suspicious actions and stop an ongoing workflow?

  • Has the complete application been tested against indirect prompt injection?

  • Is there a documented process for investigating incidents and revoking access?

  • Are AI-specific security advisories monitored alongside traditional vulnerability feeds?

An organization does not need to wait for a confirmed attack to evaluate these controls. Reviewing permissions and testing failure scenarios before deployment can reduce the likelihood and impact of an incident.

What Organizations Should Expect Next

As AI agents become more integrated with business software, the security conversation will increasingly focus on the relationship between model behavior and application permissions.

Model providers can improve resistance to malicious instructions, but application developers still control many of the capabilities that determine whether a manipulated agent can cause harm.

Organizations should therefore avoid treating prompt injection as a problem that can be solved solely by choosing a newer model or writing a more restrictive system prompt.

A more durable approach combines model-level safeguards with independent authorization, constrained tools, isolation, monitoring, and human oversight.

Security reporting will also need to distinguish between attempted attacks, observed malicious content, demonstrated vulnerabilities, and confirmed compromises. These categories provide different kinds of evidence and should not be treated as interchangeable.

The findings published in 2026 provide a reason to strengthen these practices now, especially for systems that process untrusted content while holding access to sensitive information.

Frequently Asked Questions

What is indirect prompt injection in AI?

Indirect prompt injection is an attack in which malicious instructions are embedded in external content that an AI system processes. The content may come from a webpage, email, document, or another source. If the system follows those instructions inappropriately, it may depart from its intended task or attempt an unauthorized action.

Is indirect prompt injection a real-world security concern in 2026?

Yes. Google's security research reported a 32% relative increase in detected malicious indirect prompt injection content between November 2025 and February 2026. The Cloud Security Alliance also published a research note documenting activity and reported cases involving AI agents. These findings support treating the threat seriously, although they do not establish that every detected attempt succeeded.

Can prompt injection steal private information?

Prompt injection does not automatically grant access to private information. The outcome depends on the system's permissions, connected tools, safeguards, and the information available to the agent. Risk increases when an agent can access sensitive data and communicate externally without adequate controls.

Can a system prompt completely prevent prompt injection?

No system prompt can guarantee complete protection against indirect prompt injection. Instructions can help guide model behavior, but applications should also enforce permissions, restrict tools, validate actions, and require approval for sensitive operations.

What is the most important defense against AI agent attacks?

There is no universally sufficient single defense. A strong starting point is least-privilege access combined with independent authorization and human approval for sensitive actions. These controls limit the damage an agent can cause even if its behavior is manipulated.

Why is human approval important for AI agents?

Human approval provides an additional checkpoint before consequential operations take place. It is particularly valuable for actions such as sending external communications, deleting data, changing access permissions, or making payments. Approval workflows should be combined with technical restrictions rather than used as the only security measure.

Why should organizations monitor more than CVE databases?

CVE databases primarily track identified vulnerabilities that receive CVE records. AI security incidents may also involve unsafe system design, excessive permissions, prompt injection, or misconfiguration without a corresponding CVE. Monitoring vendor advisories and security research provides a broader view of the threat landscape.

Conclusion: Secure the Agent's Capabilities, Not Just Its Prompts

Indirect prompt injection highlights a fundamental challenge in building AI systems that interact with the outside world: content the model reads can influence what it attempts to do.

Research published in 2026, including Google's findings and the Cloud Security Alliance's analysis, provides evidence that malicious prompt injection activity deserves practical security attention.

The appropriate response is not to abandon AI agents. It is to deploy them with carefully limited permissions, independent authorization checks, controlled tool access, meaningful monitoring, and human oversight for consequential actions.

The most important security principle is straightforward: an AI agent should never gain additional authority simply because untrusted content tells it to do something.

As AI systems gain more autonomy, enforcing that boundary will become an essential part of responsible AI deployment.

Editorial note: Reported detection increases, identified malicious payloads, demonstrated vulnerabilities, and confirmed successful compromises represent different types of evidence. This article distinguishes them to avoid overstating what the research establishes.