Granted Patent B2

US 12,596,839 · App. 18/787,889 · Granted Apr 7, 2026

Inventors: Tanner Burns (Austin, TX); Kwesi Cappel (Austin, TX); Kenneth Yeung (Ottawa, CA)

Assignee: HiddenLayer, Inc.

G06F21/6245G06F40/166G06F40/284

Abstract

Claims (50)

1 . A computer-implemented method comprising:

receiving an output of a generative artificial intelligence (GenAI) model responsive to a prompt by a requestor, the output being intercepted by a proxy and transmitted to a remote computing device;

tokenizing the output to result in a plurality of tokens;

determining, using the tokens, that the output comprises at least one string comprising personally identifiable information (PII), the determining using pattern recognition to identify tokens and sequence of tokens indicative of PII;

initiating at least one remediation action responsive to the determination comprising:

assigning, using a machine learning-based classifier, a PII type to each string in the output comprising PII;

determining that at least one of the PII types in the output requires redaction; and

redacting the strings having a PII type determined to require redaction to result in a modified output; and

causing the modified output to be transmitted to the requestor.

2 . The method of claim 1 , wherein the determination that least one of the PII types in the output requires redaction is based on a policy associated with the requestor.

3 . The method of claim 2 , wherein the policy is unique to the requestor and is one of a plurality of different available policies.

4 . The method of claim 3 , wherein the policy is one of a plurality of different available policies and is based on a class of users which includes the requestor.

5 . The method of claim 1 further comprising:

determining that at least one of the PII types in the output does not requires redaction;

wherein strings corresponding to the PII types not requiring redaction are not modified as part of the modified output.

6 . The method of claim 1 , wherein the GenAI model comprises a large language model.

7 . The method of claim 1 , wherein the proxy is executing further comprising:

intercepting, by a proxy executing in a computing environment in which the GenAI model executes, the output prior to the tokenizing.

8 . A system comprising:

a model computing environment comprising at least one computing device executing a generative artificial intelligence (GenAI) model and a proxy;

a monitoring computing environment comprising at least one computing device for monitoring the GenAI model by way of the proxy;

wherein a combination of the model computing environment and the monitoring environment execute operations comprising:

receiving, by the monitoring computing environment over a network from the proxy, an output of a generative artificial intelligence (GenAI) model responsive to a prompt by a requestor;

tokenizing the output to result in a plurality of tokens;

determining, using the tokens, that the output comprises at least one string comprising personally identifiable information (PII), the determining using pattern recognition to identify tokens and sequence of tokens indicative of PII;

assigning, using a classifier, a PII type to each string in the output comprising PII;

determining that at least one of the PII types in the output requires redaction;

redacting the strings having a PII type determined to require redaction to result in a modified output; and

causing the modified output to be transmitted to the requestor.

9 . The system of claim 8 , wherein the determination that least one of the PII types in the output requires redaction is based on a policy associated with the requestor.

10 . The system of claim 9 , wherein the policy is unique to the requestor and is one of a plurality of different available policies.

11 . The system of claim 10 , wherein the policy is one of a plurality of different available policies and is based on a class of users which includes the requestor.

12 . The system of claim 8 , wherein the operations further comprise:

determining that at least one of the PII types in the output does not requires redaction;

wherein strings corresponding to the PII types not requiring redaction are not modified as part of the modified output.

13 . The system of claim 8 , wherein the classifier comprises at least one machine learning model.

14 . The system of claim 8 , wherein the GenAI model comprises a large language model.

15 . A computer-implemented method comprising:

intercepting receiving a prompt from a requestor for ingestion by an artificial intelligence (AI) model executing in a model environment, the model environment comprising one or more computing devices;

determining, within a monitoring environment separate from the model environment and comprising one or more computing devices, whether the prompt comprises personally identifiable information (PII);

blocking the prompt for ingestion by the AI model when it is determined that the prompt comprises PII;

receiving, an output of the AI model responsive to the prompt by the monitoring environment, when it is determined that the prompt does not comprise PII;

determining, within the monitoring environment, whether the output comprises PII;

allowing the output to be transmitted to the requestor from the model environment if it is determined that the output does not comprise PII; and

selectively redacting the output before transmission from the model environment to the requestor based on a policy associated the requestor, the policy specifying levels of PII that require redaction for the requestor.

16 . The method of claim 15 , wherein a PII type is assigned to each string in the output comprising PII.

17 . The method of claim 16 , wherein the PII type is assigned by a machine learning-based classifier.

18 . The method of claim 15 , wherein the AI model is a large language model.

19 . The method of claim 15 , wherein the selective redaction is based on a policy associated with the requestor.

20 . The method of claim 19 , wherein the policy is one of a plurality of different available policies and is based on a class of users which includes the requestor.

Assignments (1)

Continuity (2)

References Cited (159)