# US 2026/0087131 A1 · Large language model generated endpoint detection and response system

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0087131 A1
- Document ID: US20260087131A1
- Status: Published application
- Kind: A1
- Application number: 18/963,406
- Filed: 2024-11-27
- Published: 2026-03-26
- Earliest claimed: 2024-09-25
- Assignees: Microsoft Technology Licensing, LLC
- Applicants: Microsoft Technology Licensing, LLC
- Inventors: Zheng DONG; Everton Luís BERZ; Jie LI; Patrick ESTAVILLO; Geoffrey Lyall MCDONALD; Dmitri BEKERMAN; Evan John ARGYLE; Moukhik MISRA
- Web page: https://patentgrep.com/patent/US20260087131A1

## CPC

- G06F 21/554: {involving event detection and direct action}
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

The disclosed configurations generate synthetic attack data that emulates the markers of a cyberattack. The generated synthetic attack data is used to train an attack detection machine learning model that detects and mitigates actual cyberattacks in real-time. Synthetic attack data is generated by a synthetic attack data generation model, which is trained with a synthetic attack data generation prompt. The synthetic attack data generation prompt is constructed out of attack data samples and a prompt guideline. The prompt guideline is created from attack procedure descriptions, such as security blog posts or other write-ups about actual cyberattacks. Prompt guidelines may include samples of actual attack data that indicate how to format synthetic attack data. Once deployed, the attack detection machine learning model infers the occurrence of a cyberattack from log entries. Detected cyberattacks may be mitigated in an automated or semi-automated manner.

## Claims

### Claim 1

A method comprising:
receiving an attack procedure description;
generating a prompt guideline from the attack procedure description;
generating a synthetic attack data training prompt from the prompt guideline and an attack data sample;
generating, with the synthetic attack data training prompt, synthetic attack data;
training an attack detection machine learning model with the synthetic attack data;
identifying a security incident with the attack detection machine learning model; and
mitigating the security incident.

### Claim 2 (depends on 1)

The method of claim 1, wherein the attack detection machine learning model is trained with benign action data that indicates operations taken on a computing device during a benign interaction.

### Claim 3 (depends on 1)

The method of claim 1, wherein generating synthetic attack data comprises generating log entries that mimic real-world attack data.

### Claim 4 (depends on 1)

The method of claim 1, wherein training the attack detection machine learning model comprises refining an existing large language model.

### Claim 5 (depends on 1)

The method of claim 1, wherein the prompt guideline comprises samples of actual attack data.

### Claim 6 (depends on 5)

The method of claim 5, wherein the synthetic attack data is formatted based on the samples of actual attack data.

### Claim 7 (depends on 1)

The method of claim 1, wherein the prompt guideline is generated in part with a prompt guideline generation model.

### Claim 8

A computer-readable storage medium having computer-executable instructions stored thereupon that, when executed by a processing system, cause the processing system to:
receive an attack procedure description;
generate a synthetic attack data training prompt from the attack procedure description and an attack data sample;
generate, with the synthetic attack data training prompt, synthetic attack data;
train an attack detection machine learning model with the synthetic attack data;
identify a security incident with the attack detection machine learning model; and
mitigate the identified security incident.

### Claim 9 (depends on 8)

The computer-readable storage medium of claim 8, wherein the synthetic attack data training prompt is generated based on a prompt guideline that is derived from the attack data sample.

### Claim 10 (depends on 8)

The computer-readable storage medium of claim 8, wherein the prompt guideline includes samples of actual attack data.

### Claim 11 (depends on 8)

The computer-readable storage medium of claim 8, the attack data sample comprises an event from an event log.

### Claim 12 (depends on 8)

The computer-readable storage medium of claim 8, wherein the attack detection machine learning model is trained on real attack data.

### Claim 13 (depends on 8)

The computer-readable storage medium of claim 8, wherein the attack detection machine learning model is trained on benign action data.

### Claim 14

A system comprising:
a processor;
a memory storing instructions that, when executed by the processor, cause the system to perform operations comprising:
receiving an attack procedure description;
generating a prompt guideline from the attack procedure description;
generating a synthetic attack data training prompt from the prompt guideline and an attack data sample;
generating, with the synthetic attack data training prompt, synthetic attack data;
training an attack detection machine learning model with the synthetic attack data;
identifying a security incident with the attack detection machine learning model; and
mitigating the security incident.

### Claim 15 (depends on 14)

The system of claim 14, wherein the prompt guideline is generated by a prompt guideline generation model.

### Claim 16 (depends on 15)

The system of claim 15, wherein the prompt guideline generation model infers the prompt guideline from a prompt generation guideline prompt and the attack procedure description.

### Claim 17 (depends on 14)

The system of claim 14, wherein the synthetic attack data is inferred from the synthetic attack data training prompt by a synthetic attack data generation model.

### Claim 18 (depends on 14)

The system of claim 14, wherein the security incident is provisionally identified by identifying a log entry with a low-parameter version of the attack detection model.

### Claim 19 (depends on 18)

The system of claim 18, wherein a refined version of the attack detection model confirms that the provisionally identified security incident is an actual security incident.

### Claim 20 (depends on 14)

The system of claim 14, wherein the security incident is mitigated by providing the alert to a large language model as part of a prompt for a mitigation procedure and executing the mitigation procedure.
