MITRE ATT&CK technique · T1699
Direct Prompt Injection
What is direct prompt injection? Direct prompt injection is an adversary supplying instructions directly to an AI system through its normal input, attempting to override its intended behaviour — for example to reveal instructions, bypass content policies or misuse tools.
For leadership
Executive takeaway
How it works in AWS
- The adversary crafts input that conflicts with or overrides system instructions.
- Techniques include role-play, encoding, and multi-turn manipulation.
- If the model has tools, the adversary tries to trigger them with chosen parameters.
What the attacker needs
- Access to the AI system’s input
Relevant AWS API calls
| API / event | Why it matters |
|---|---|
| bedrock:InvokeModel / Converse | Model invocation where the input arrives. |
A risky permission is not an observed attack
Public-facing assistants with tools that act or access data beyond the user’s rights.
Guardrail blocks, unusual tool parameters, attempts to extract system instructions.
Preventive and limiting controls
- Authorization enforced by tools
- Guardrails / input filters
- Rate limiting and abuse monitoring
Mapped cloud controls
- restrictiveInput and output guardrails (defense in depth)
Prompt-attack filters reduce the success rate of direct jailbreak and injection attempts.
- preventiveAuthorization enforced by tools, not the model
Even if a user talks the model into an action, the tool should refuse what that user is not allowed to do.
See assumptions and validation evidence in the Cloud Control Explorer →
Detection signals and telemetry
- Prompt-attack filter hits — Guardrail metrics / logs
- Tool calls rejected by authorization — Tool audit logs
MITRE detection strategies (official, cloud-relevant analytics)
- DET0941 Instruction Override with Unauthorized AI Activity
- AN2292 (SaaS): An analyst observes a user-originated instruction submitted to an AI application that attempts to override protected instructions or operational constraints. Within the same authenticated interaction, the AI system subsequently attempts a tool invocation, resource access, or workflow transition inconsistent with the authorized task or capability scope.Log sources: m365:unified — CopilotInteraction
MITRE mitigations
Official MITRE description (excerpt)
Adversaries may input malicious prompts directly into an LLM’s user interface. The prompts can be designed to bypass safeguards, manipulate context, override previous instructions, or coerce the LLM into unauthorized behavior.
Adversaries may use direct prompt injection to Generate Content prohibited by the LLM's guardrails or to induce the system to utilize its capabilities in unintended or prohibited ways. Adversaries may achieve this by instructing the LLM to ignore previous instructions or safety rules, or by impersonating a role such as a researcher or privileged user.
Common misconceptions
- “Guardrails make injection impossible.”
- Filters are probabilistic and bypassable; they are one layer.
AI security
Related MITRE ATLAS techniques
ATLAS and ATT&CK are separate frameworks. These relationships are listed explicitly, with their basis.
- AML.T0051.000 LLM Prompt Injection: DirectRelated behavior (Securitain analysis) — Both describe an adversary supplying malicious instructions directly to an AI system’s input.
- AML.T0054 LLM JailbreakRelated AI-specific behavior (Securitain analysis).
Attack scenarios using this technique
No scenario in the library uses this technique yet.
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