# US 12,111,926 B1 · Generative artificial intelligence model output obfuscation

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

- Number: US 12,111,926 B1
- Document ID: US12111926B1
- Status: Granted
- Kind: B1
- Application number: 18/669,379
- Filed: 2024-05-20
- Granted: 2024-10-08
- Earliest claimed: 2024-05-20
- Assignees: HiddenLayer, Inc.
- Applicants: HiddenLayer, Inc.
- Inventors: David Beveridge; Tanner Burns; Kwesi Cappel; Kenneth Yeung
- Web page: https://patentgrep.com/patent/US12111926B1

## CPC

- G06F 21/554: {involving event detection and direct action}
- G06F 2221/033: Test or assess software
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

An analysis engine receives data characterizing a prompt for ingestion by a generative artificial intelligence (GenAI) model. The analysis engine, using a determines using, for example, a classifier or blocklist, that the prompt comprises or is indicative of malicious content or otherwise elicits undesired model behavior. Similarly, outputs of the GenAI model can be analyzed to determine whether they comprise malicious content or cause the model to behave in an undesired manner. The output is inputted into a GenAI model along with obfuscation instructions to generate an output which is returned to the requesting user. Related apparatus, systems, techniques and articles are also described.

## Claims

### Claim 1

A computer-implemented method comprising:
receiving, from each of a plurality of requesters, data characterizing a corresponding prompt for ingestion by a first generative artificial intelligence (GenAI) model;
determining, for each prompt, that the prompt comprises malicious content or elicits undesired model behavior; and
initiating, in response to the determining, at least one remediation action;
wherein, for a first subset of the prompts, the at least one remediation action comprises:
inputting at least a portion of the received data into the first GenAI model to obtain a first output;
inputting at least a portion of the first output along with obfuscation instructions into a second, different GenAI model to obtain a second output; and
returning data characterizing the second output to the requester;
wherein, for a second subset of the prompts, the at least one remediation action comprises:
blocking an Internet Protocol (IP) and/or medium access control address (MAC) address of a corresponding requester from accessing the first GenAI model.

### Claim 2 (depends on 1)

The method of claim 1, wherein the determination is based on a classification by a classifier.

### Claim 3 (depends on 2)

The method of claim 2, wherein the determination is based on a blocklist defining content deemed to be malicious or eliciting undesired model behavior.

### Claim 4 (depends on 1)

The method of claim 1, wherein at least one of the first GenAI model or the second GenAI model comprises a large language model.

### Claim 5 (depends on 1)

The method of claim 1, wherein the obfuscation instructions identify a portion of content in the first output which requires modification.

### Claim 6 (depends on 4)

The method of claim 4, wherein the modification comprises generating synthetic data corresponding to the identified portion of content.

### Claim 7 (depends on 5)

The method of claim 5, wherein the modification comprises redacting data corresponding to the identified portion of content.

### Claim 8 (depends on 1)

The method of claim 1 further comprising inputting at least a portion of the received data into the second GenAI model along with the at least a portion of the first output and the obfuscation instructions to obtain the second output.

### Claim 9

A computer-implemented method comprising:
receiving, from each of a plurality of requesters, data characterizing a corresponding prompt for ingestion by a first generative artificial intelligence (GenAI) model;
inputting, for each prompt at least a portion of the received data into the first GenAI model to obtain a first output;
determining, for each first output, that the first output comprises or elicits malicious or undesired content; and
initiating, in response to the determining, at least one remediation action;
wherein, for a first subset of the prompts, the at least one remediation action comprises:
inputting at least a portion of the first output along with obfuscation instructions to a second, different GenAI model to obtain a second output; and
returning data characterizing the second output to the requester;
wherein, for a second subset of the prompts, the at least one remediation action comprises:
blocking an Internet Protocol (IP) and/or medium access control address (MAC) address of a corresponding requester from accessing the first GenAI model.

### Claim 10 (depends on 9)

The method of claim 9, wherein the determination is based on a classification by a classifier.

### Claim 11 (depends on 9)

The method of claim 9, wherein the determination is based on a blocklist defining content deemed to be malicious or eliciting undesired model behavior.

### Claim 12 (depends on 9)

The method of claim 9, wherein at least one of the first GenAI model or the second GenAI model comprises a large language model.

### Claim 13 (depends on 9)

The method of claim 9, wherein the obfuscation instructions identify a portion of content in the first output which requires modification.

### Claim 14 (depends on 13)

The method of claim 13, wherein the modification comprises generating synthetic data corresponding to the identified portion of content.

### Claim 15 (depends on 13)

The method of claim 13, wherein the modification comprises redacting data corresponding to the identified portion of content.

### Claim 16 (depends on 1)

The method of claim 1 further comprising inputting at least a portion of the received data into the second GenAI model along with the at least a portion of the first output and the obfuscation instructions to obtain the second output.

### Claim 17

A computer-implemented method comprising:
receiving, from each of a plurality of requesters, a corresponding prompt for ingestion by a first generative artificial intelligence (GenAI) model;
determining, for each prompt, that the prompt comprises or elicits malicious content or undesired model behavior;
initiating, in response to the determining, at least one remediation action;
wherein, for a first subset of the prompts, the at least one remediation action comprises:
inputting the prompt into the first GenAI model to obtain a first output;
inputting the prompt, the first output, and obfuscation instructions into a second, different GenAI model to obtain a second output; and
returning data characterizing the second output to the corresponding requester;
wherein, for a second subset of the prompts, the at least one remediation action comprises:
blocking an Internet Protocol (IP) and/or medium access control address (MAC) address of a corresponding requester from accessing the first GenAI model.

### Claim 18

A computer-implemented method comprising:
receiving, from each of a plurality of requesters, a prompt for ingestion by a first generative artificial intelligence (GenAI) model;
inputting, for each prompt, the prompt into the first GenAI model to obtain a first output;
determining, for each first output, that the first output comprises or elicits malicious or undesired model behavior;
initiating, in response to the determining, at least one remediation action;
wherein, for a first subset of the prompts, the at least one remediation action comprises:
inputting the prompt along with obfuscation instructions into a second GenAI model to obtain a second output; and
returning data characterizing the second output to the requester;
wherein, for a second subset of the prompts, the at least one remediation action comprises:
blocking an Internet Protocol (IP) and/or medium access control address (MAC) address of a corresponding requester from accessing the first GenAI model.

### Claim 19 (depends on 17)

The method of claim 17, wherein the determination is based on a classification by a classifier.

### Claim 20 (depends on 17)

The method of claim 17, wherein the determination is based on a blocklist defining content deemed to be malicious or eliciting undesired model behavior.

### Claim 21 (depends on 18)

The method of claim 18, wherein the determination is based on a blocklist defining content deemed to be malicious or eliciting undesired model behavior.

### Claim 22 (depends on 18)

The method of claim 18, wherein the determination is based on a classification by a classifier.
