T1699: Direct Prompt Injection
Adversaries may input malicious prompts directly into an LLM’s user interface.[1] The prompts can be designed to bypass safeguards, manipulate context, override previous instructions, or coerce the LLM into unauthorized behavior.[2]
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.[3][4] 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.[5]
Security context for executives and security teams
T1699: Direct Prompt Injection describes Adversaries may input malicious prompts directly into an LLM’s user interface.(Citation: NIST Direct Prompt Injection) The prompts can be designed to bypass safeguards, manipulate context, override previous instructions, or coerce the LLM into unauthorized behavior.(Citation: OWASP Prompt Injection) Adversaries may use direct prompt injection to [Generate Content](https://attack.mitre.org/techniques/T1683) prohibited by the LLM's guardrails or to induce the system to utilize its capabilities in unintended or prohib...
Executive priority
T1699: Direct Prompt Injection is an official MITRE ATT&CK technique. Glexia treats it as defensive behavior context for prioritizing monitoring, control validation, and response planning without using the object by itself as an attribution claim.
Technical view
Security teams should validate T1699: Direct Prompt Injection by reviewing the official ATT&CK relationships, mapped tactics (execution), supported platforms (Windows, SaaS), and available local telemetry before making detection or mitigation decisions.
Likely telemetry
- Official ATT&CK relationships and object metadata
- Endpoint process, command-line, and script execution logs
- Network, endpoint, and security-tool telemetry
Detection direction
- Validate whether T1699: Direct Prompt Injection appears in your detection coverage and tabletop scenarios.
- Use the object to align executive risk language with SOC, incident response, and detection engineering work.
- Do not treat ATT&CK relationship context as attribution without corroborating evidence.
Mitigation priorities
- Map the object to existing controls and identify missing telemetry or response ownership.
- Prioritize mitigations that reduce exposure on the listed platforms and tactics.
- Review adjacent ATT&CK relationships before changing policy, detections, or reporting language.
Additional notes and limits
Baseline Glexia take generated from the official MITRE ATT&CK STIX object, source hash, tactics, platforms, and detection fields. It is safe to replace with a richer model-generated take for the same source hash later.
This baseline take is source-grounded and schema-validated, but it does not include environment-specific telemetry, incident evidence, or threat-intelligence corroboration.
Generated from the cited source records. This long-tail analysis has not been individually reviewed by a named human.
Direct Prompt Injection
Adversaries may input malicious prompts directly into an LLM’s user interface.[1] The prompts can be designed to bypass safeguards, manipulate context, override previous instructions, or coerce the LLM into unauthorized behavior.[2]
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.[3][4] 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.[5]
How security teams should use this page
Treat this object as behavior context, not an attribution claim. Validate the related groups, software, data sources, and mitigations against official ATT&CK relationships and your own telemetry before making control-coverage decisions.
Groups, software, and campaigns
C0062: Anthropic AI-orchestrated Campaign
The Anthropic AI-orchestrated Campaign was conducted in September 2025 by a likely China nexus espionage actor identified as GTG-1002. The Anthropic AI-orchestrated Campaign was a highly coordinated operation that manipulated Claude Code to perform reconnaissance, vulnerability discovery, exploitation, lateral movement, credential harvesting, data analysis, and exfiltration operations at approximately 30 entities in the technology, financial, chemical, and government sectors. During the Anthropic AI-orchestrated Campaign, human operators used Claude Code agents and Model Context Protocol (MCP) tools to automate cyber operations. Operators broke attacks into discrete tasks, used crafted prompts, and established personas to bypass AI guardrails, enabling the agents to execute the operations with minimal human involvement.[1][2]
C0064: AI-Assisted Breach of Mexico's Government Infrastructure
The AI-Assisted Breach of Mexico's Government Infrastructure is a campaign where human operators used Claude Code AI and OpenAI GPT GPT-4.1 AI to perform reconnaissance, vulnerability exploitation, privilege escalation, discovery, credential harvesting, and exfiltration operations. The attack impacted several Mexican government institutions including the Federal Tax Authority (Servicio de Administracion Tributaria), Mexico State Government (Estado de Mexico), Mexico City Civil Registry (Registro Civil de CDMX), Jalisco State Government, National Electoral Institute (Instituto Nacional Electoral), Michoacan State Government, SADM Monterrey Municipal Water Utility (Agua y Drenaje), Tamaulipas State Government, and the Mexico City Health Department (Salud CDMX). The attack resulted in the theft of large amounts of sensitive government data and civilian records. Please note that due to the breadth of this campaign, public reporting only provides select details and as a result this campaign should not be considered comprehensive.[1]
All related ATT&CK context
Mitigation direction
Object version and sync metadata
The fields below describe the current mirrored snapshot. When Glexia retains multiple ATT&CK source imports, you can open the table to compare the same object across releases (hashes and MITRE timestamps). For MITRE’s own release notes and roadmap, see ATT&CK resources — Updates.
Imported snapshots across ATT&CK releases(1)
| Release | Bundle imported | Object version | Modified | Status | Raw hash |
|---|---|---|---|---|---|
| 20.0 | 1.0 | Current bundle | 0ce712980649… |
Mirrored ATT&CK source object
The raw object is retained through the mirrored ATT&CK source bundle and object hash. The raw endpoint returns the exact object from the mirrored bundle when available.
External references and citations
MITRE external references are preserved separately from Glexia analysis so citations remain traceable to their original source records.
- [1]NIST Direct Prompt Injection
NIST. (n.d.). direct prompt injection. Retrieved September 18, 2026.
Open source URL - [2]OWASP Prompt Injection
OWASP. (n.d.). LLM01:2025 Prompt Injection. Retrieved September 18, 2026.
Open source URL - [3]Gemini AI Misuse Jan 2025
Google Threat Intelligence Group. (2025, January). Adversarial Misuse of Generative AI. Retrieved September 18, 2026.
Open source URL - [4]Palo Alto Prompt Injection
Palo Alto Networks. (n.d.). What Is a Prompt Injection Attack? [Examples & Prevention]. Retrieved September 18, 2026.
Open source URL - [5]Cybersec Guru Prompt Injection
The CyberSec Guru. (2026, June 7). The Meta AI exploit: how a prompt injection flaw bypassed 2FA to steal million-dollar Instagram accounts. Retrieved September 18, 2026.
Open source URL - [6]mitre-attackT1699Open source URL
Source: MITRE ATT&CK®. © 2026 The MITRE Corporation. This work is reproduced and distributed with the permission of The MITRE Corporation. MITRE ATT&CK and ATT&CK are registered trademarks of The MITRE Corporation. Glexia is not affiliated with or endorsed by MITRE.
