IP Library › Granted Patent US 12,596,789
Granted Patent B2
US 12,596,789 · App. 18/510,382 · Granted Apr 7, 2026

Large language model (LLM) interaction security sandbox

Inventors: Hiranmayi Palanki (Tampa, FL); Shankar Djeyassilane (Markham, CA)
Assignee: American Express Travel Related Services Company
G06F21/53G06F40/20
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,596,789
App. No.
18/510,382
Granted
Apr 7, 2026
Kind
B2
Abstract

Disclosed are various approaches for large language model (LLM) interaction security sandboxing. A client device can execute an LLM security sandbox that includes at least one LLM communications sanitization process. The LLM security sandbox can perform the at least one LLM communications sanitization process on the LLM message to generate an approved LLM message. The client device can provide access to the approved LLM message by at least generating a user interface that includes the approved LLM message, or transmitting the approved LLM message from the client device to the LLM service.

Claims (40)

1 . A system, comprising:

a client device comprising at least one processor and at least one memory; and

machine-readable instructions stored in the at least one memory that, when executed by the at least one processor, cause the client device to at least:

create, in the client device, a Large Language Model (LLM) security sandbox comprising at least one LLM communications sanitization process;

provide an LLM message to the LLM security sandbox, wherein the LLM message is generated or received for communications over a network with an LLM service;

receive, from the LLM security sandbox, an approved LLM message generated based at least in part on performing the at least one LLM communications sanitization process on the LLM message; and

provide access to the approved LLM message by at least:

generating a user interface that includes the approved LLM message, or transmitting the approved LLM message to the LLM service over the network.

2 . The system of claim 1 , wherein the approved LLM message is a modified version of the LLM message.

3 . The system of claim 2 , wherein the modified version of the LLM message is generated using a moderator LLM in the LLM security sandbox.

4 . The system of claim 1 , wherein the LLM security sandbox comprises an LLM virtual Document Object Model (DOM) that is a virtual representation of an LLM DOM for a web page or a web application.

5 . The system of claim 1 , wherein the LLM message is an LLM input message entered through a user prompt.

6 . The system of claim 1 , wherein the LLM message is an LLM response received from the LLM service based at least in part on an LLM input message.

7 . The system of claim 1 , wherein the approved LLM message is provided with a moderator comment that indicates a result of the at least one LLM communications sanitization process.

8 . A method, comprising:

executing, by a client device, a Large Language Model (LLM) security sandbox, wherein the LLM security sandbox comprises at least one LLM communications sanitization process;

identifying, by the client device, an LLM message generated or received for communications with an LLM service over a network;

performing, by the client device, the at least one LLM communications sanitization process on the LLM message to generate an approved LLM message; and

providing, by the client device, access to the approved LLM message by at least:

generating a user interface that includes the approved LLM message, or transmitting the approved LLM message from the client device to the LLM service over the network.

9 . The method of claim 8 , further comprising:

generating, by the client device, a moderator comment that indicates a result of the at least one LLM communications sanitization process.

10 . The method of claim 8 , wherein the approved LLM message is a modified version of the LLM message.

11 . The method of claim 10 , wherein the modified version of the LLM message is generated using a moderator LLM in the LLM security sandbox.

12 . The method of claim 8 , wherein the LLM security sandbox comprises an LLM virtual Document Object Model (DOM) that is a virtual representation of an LLM DOM for a web page or a web application.

13 . The method of claim 8 , wherein the LLM message is an LLM input message entered through a user prompt.

14 . The method of claim 8 , wherein the LLM message is an LLM response received from the LLM service based at least in part on an LLM input message.

15 . A system, comprising:

at least one computing device comprising at least one processor and at least one memory; and

machine-readable instructions stored in the at least one memory that, when executed by the at least one processor, cause the at least one computing device to at least:

execute a Large Language Model (LLM) security sandbox using a client device, wherein the LLM security sandbox comprises at least one LLM communications sanitization process;

identify an LLM message generated or received for communications with an LLM service over a network;

perform the at least one LLM communications sanitization process on the LLM message to generate an approved LLM message; and

provide access to the approved LLM message by at least:

generating a user interface that includes the approved LLM message, or transmitting the approved LLM message from the client device to the LLM service over the network.

16 . The system of claim 15 , wherein the approved LLM message is a modified version of the LLM message.

17 . The system of claim 16 , wherein the modified version of the LLM message is generated using a moderator LLM in the LLM security sandbox.

18 . The system of claim 15 , wherein the LLM security sandbox comprises an LLM virtual Document Object Model (DOM) that is a virtual representation of an LLM DOM for a web page or a web application.

19 . The system of claim 15 , wherein the LLM message is an LLM input message entered through a user prompt.

20 . The system of claim 15 , wherein the LLM message is an LLM response received from the LLM service based at least in part on an LLM input message.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 20, 2023
From: PALANKI, HIRANMAYI; DJEYASSILANE, SHANKAR
To: AMERICAN EXPRESS TRAVEL RELATED SERVICES COMPANY, INC.
Reel/Frame 065618/0307 →
Continuity (1)
Related Publication 20250156527A1 · May 15, 2025
References Cited (5)
US 20240193374A1 · Wolfe · 2024 [cited by examiner]
US 20240388551A1 · AlFardan · 2024 [cited by examiner]
US 20240427994A1 · Odland · 2024 [cited by examiner]
US 20250039219A1 · Mirkis · 2025 [cited by examiner]
US 20250103746A1 · Gharibi · 2025 [cited by examiner]