MITRE ATT&CK technique · T1698

Indirect Prompt Injection

What is indirect prompt injection? Indirect prompt injection places adversary instructions in content an AI system ingests during normal operation — retrieved documents, emails, web pages, tool output — rather than in the user’s own message. The model may treat that content as instructions.

ExecutionPlatforms: SaaS, WindowsATT&CK v20.1 · source modified · View on MITRE ATT&CK

For leadership

Executive takeaway

There is no complete fix today. Design AI systems so that following an injected instruction cannot cause serious harm: least privilege, human approval and authorization outside the model.

How it works in AWS

  1. The adversary places crafted text where the AI system will read it.
  2. The system retrieves or receives that content as context.
  3. The model follows the embedded instructions, possibly invoking tools or emitting data.

What the attacker needs

  • The AI system ingests content the adversary can influence

Relevant AWS API calls

AWS API calls relevant to T1698
API / eventWhy it matters
bedrock:Retrieve / RetrieveAndGenerateKnowledge-base retrieval that may surface poisoned content.
bedrock:InvokeAgentAgents that may act on injected instructions.

A risky permission is not an observed attack

Potential exposure

Agents that read untrusted content and also hold write tools or broad roles.

Observed behavior

Tool calls or outputs that follow instructions present in retrieved content.

Preventive and limiting controls

  • Treat retrieved content as untrusted; restrict tools after reading it
  • Authorization enforced by tools
  • Human approval for consequential actions
  • Guardrails as defense in depth

Mapped cloud controls

  • restrictiveTreat retrieved and tool-returned content as untrusted

    Indirect prompt injection depends on the agent acting on instructions hidden in content it retrieves; separating and constraining untrusted content limits that influence.

  • preventiveAuthorization enforced by tools, not the model

    If tools enforce the end user’s permissions, injected instructions cannot make the agent access data the user could not.

  • restrictiveHuman approval for consequential agent actions

    Human confirmation of consequential actions interrupts injected instructions before they take effect.

See assumptions and validation evidence in the Cloud Control Explorer →

Detection signals and telemetry

  • Instruction-like content in ingestion pipelines — Ingestion scanning
  • Tool calls inconsistent with user intent — Agent audit logs

MITRE detection strategies (official, cloud-relevant analytics)

  • DET0942 Retrieved Instruction with Unauthorized Agent Action
    • AN2293 (SaaS): An analyst observes an AI application retrieve or ingest lower-trust content containing instructions that attempt to redirect the application's behavior. Within the same agent execution context, the AI system subsequently attempts a tool invocation, resource access, or workflow transition outside the actions authorized for the original task or retrieved resource.Log sources: AWS:bedrock — AgentTrace.knowledgeBaseLookupInput; AWS:bedrock — AgentTrace.actionGroupInvocationInput

MITRE mitigations

Official MITRE description (excerpt)

Adversaries may deliver malicious prompts to an LLM indirectly by embedding the prompts in content that the adversary expects may be processed by an LLM. These prompts can be included in web pages, documents, emails, tickets, code, or other sources that an agent, copilot, or other AI system may retrieve and process. When an LLM processes this content, the embedded instructions may cause the system to perform unintended behaviors.

Adversaries may use indirect prompt injection to Impair Autonomous Defenses, and for unauthorized execution of any capabilities available to the LLM.

Common misconceptions

“A better system prompt prevents injection.”
System prompts are not a security boundary; injected text competes with them.

AI security

Related MITRE ATLAS techniques

ATLAS and ATT&CK are separate frameworks. These relationships are listed explicitly, with their basis.

  • AML.T0051.001 LLM Prompt Injection: IndirectRelated behavior (Securitain analysis) — Both describe instructions delivered through content an AI system ingests rather than through the user. ATLAS models it as an LLM prompt-injection sub-technique; Enterprise ATT&CK models it as an Execution technique.
  • AML.T0070 RAG PoisoningRelated AI-specific behavior (Securitain analysis).

Direct and indirect prompt injection →RAG, memory and retrieved-content poisoning →

Attack scenarios using this technique

References

MITRE ATT&CK® and ATT&CK® are registered trademarks of The MITRE Corporation. ATT&CK data © 2026 The MITRE Corporation. This work is reproduced and distributed with the permission of The MITRE Corporation (ATT&CK v20.1). MITRE ATLAS™ data (v2026.10) © MITRE, used under the Apache License 2.0. Securitain is not affiliated with or endorsed by MITRE. Explanations, scenarios and control mappings on this page are Securitain’s own analysis. Sources and licensing

Stay in the loop

Cloud security insights, early access & compliance updates

Weekly InsightsEarly AccessExpert Content
No spam, everUnsubscribe anytimeWeekly updates
Chat with Securitain