Granted Patent B2

US 12,572,777 · App. 18/811,610 · Granted Mar 10, 2026

Inventors: Andrew Davis (Portland, OR); Amelia Kawasaki (Corvallis, OR)

Assignee: HiddenLayer, Inc.

G06N3/0442G06N3/0475

Abstract

Claims (45)

1 . A computer-implemented method comprising:

receiving, from a model computing environment, data characterizing a multimodal prompt for ingestion by a generative artificial intelligence (GenAI) model, the GenAI model comprising a pipeline including an encoder for processing the received data prior to ingestion by a plurality of layers including at least one intermediate layer, the GenAI model further comprising a transformer architecture with residual streams;

redirecting the multimodal prompt to a quantized proxy of the GenAI model that replicates a layer topology of the GenAI model and is more compact;

capturing an intermediate result derived from the at least one intermediate layer of the quantized proxy of the GenAI model responsive to the inputting of the received data after it is processed by the encoder, wherein the intermediate result comprises residual-stream activation vectors derived from the at least one intermediate layer of the quantized proxy of the GenAI model via forward-pass hooks applied to the transformer layers;

determining, based on the intermediate result and using a classifier, whether the prompt seeks to cause the GenAI model to behave in an undesired manner, the classifier being trained using a data set generated using a policy mapping intermediate results to undesired model behavior including a plurality of enumerated prohibited actions; and

providing data characterizing the determination to a consuming application or process, the consuming application or process:

preventing the multimodal prompt from being ingested by the GenAI model in response to determining that the prompt elicits undesired actions by the GenAI model; and

allowing the multimodal prompt to be ingested by the GenAI model in response to determining that the prompt does not elicit undesired actions by the GenAI model.

2 . The method of claim 1 , wherein the GenAI model comprises a mixture of experts (MoE) model and the intermediate result comprises outputs from at least a subset of experts in the MoE model.

3 . The method of claim 1 , wherein the consuming application or process flags the prompt based on the determination.

4 . The method of claim 1 , wherein the consuming application or process blocks an internet protocol (IP) address of a requester of the prompt based on the determination.

5 . The method of claim 1 , wherein the consuming application or process causes subsequent prompts from an internet protocol (IP) address of a requester of the prompt to be modified based on the determination and causes the modified prompt to be ingested by the GenAI model.

6 . The method of claim 1 , wherein the consuming application or process prevents the prompt from being input into the GenAI model based on the determination.

7 . The method of claim 1 , wherein the consuming application or process allows the prompt to be input into the GenAI model based on the determination.

8 . The method of claim 1 , wherein the consuming application or process modifies the prompt based on the determination and causes the modified prompt to be ingested by the GenAI model.

9 . The method of claim 1 further comprising: quantizing the GenAI model prior to capturing the intermediate result.

10 . The method of claim 1 , wherein a policy maps intermediate results to one or more prohibited actions.

11 . The method of claim 10 , wherein the one or more prohibited actions include seeking one or more of: a response in a non-approved spoken or written language, computer code, sensitive information, a response to an encrypted prompt, a prompt containing two or modalities, or a prompt in a first modality obfuscating information in a second modality.

12 . The method of claim 1 , wherein the GenAI model is a state-based text analysis model.

13 . The method of claim 12 , wherein the state-based text analysis model comprises: a large language model.

14 . The method of claim 12 , wherein the state-based text analysis model comprises: a long short-term memory model (LSTM).

15 . A computer-implemented method comprising:

receiving data characterizing a multimodal prompt for ingestion by a generative artificial intelligence (GenAI) model, the GenAI model comprising a pipeline including an encoder for processing the received data prior to ingestion by a plurality of layers including at least one intermediate layer;

capturing, from a quantized proxy of the GenAI model and based on the received data after it is processed by the encoder, an intermediate result derived from the at least one intermediate layer, the quantized proxy of the GenAI model being separate from and more compact than the GenAI model, wherein the intermediate result comprises residual-stream activation vectors derived from the at least one intermediate layer of the quantized proxy of the GenAI model via forward-pass hooks applied to one or more transformer layers;

determining, using a classifier and based on the intermediate result, whether the prompt elicits undesired actions by the GenAI model, the classifier being trained using a data set generated according to a policy which maps acceptable and undesired actions to intermediate results;

providing data characterizing the determination to a consuming application or process, the consuming application or process:

preventing the multimodal prompt from being ingested by the GenAI model in response to determining that the prompt elicits undesired actions by the GenAI model; and

allowing the multimodal prompt to be ingested by the GenAI model in response to determining that the prompt does not elicit undesired actions by the GenAI model.

16 . A computer-implemented method for policy-based gating of multimodal prompts in a generative artificial intelligence (GenAI) system, the method comprising:

receiving, by a proxy computing system, data characterizing a multimodal prompt for ingestion by a production GenAI model, the GenAI model comprising a transformer architecture with a plurality of layers including at least one intermediate layer and residual streams;

redirecting the multimodal prompt to a quantized proxy model that shares a layer topology with the production GenAI model, the quantized proxy model configured to process the prompt without generating a final output;

capturing, by the proxy computing system and via forward-pass hooks applied to at least one transformer layer of the quantized proxy model, per-token activation vectors from residual streams at a predetermined intermediate layer, wherein the capturing occurs after encoding of the prompt and before any output is generated by the production GenAI model;

applying, by the proxy computing system, a classifier trained using a policy-labeled dataset that maps internal activation patterns to a plurality of enumerated prohibited actions, the prohibited actions comprising at least: (i) generation of computer code, (ii) response in a non-approved language, (iii) response to an encrypted prompt, or (iv) cross-modal obfuscation;

determining, by the proxy computing system and based on the per-token activation vectors, whether the prompt is likely to elicit a prohibited action by the production GenAI model; and

in response to the determination, gating the prompt by preventing the prompt from being ingested by the production GenAI model when the determination indicates a prohibited action, and allowing ingestion otherwise, such that the production GenAI model does not process prompts likely to elicit prohibited actions.

17 . A system comprising:

at least one data processor; and

memory storing instructions which, when executed by the at least one data processor, result in operations comprising:

receiving, from a model computing environment, data characterizing a multimodal prompt for ingestion by a generative artificial intelligence (GenAI) model, the GenAI model comprising a pipeline including an encoder for processing the received data prior to ingestion by a plurality of layers including at least one intermediate layer, the GenAI model further comprising a transformer architecture with residual streams;

redirecting the multimodal prompt to a quantized proxy of the GenAI model that replicates a layer topology of the GenAI model and is more compact;

capturing an intermediate result derived from the at least one intermediate layer of the quantized proxy of the GenAI model responsive to the inputting of the received data after it is processed by the encoder, wherein the intermediate result comprises residual-stream activation vectors derived from the at least one intermediate layer of the quantized proxy of the GenAI model via forward-pass hooks applied to the transformer layers;

determining, based on the intermediate result and using a classifier, whether the prompt seeks to cause the GenAI model to behave in an undesired manner, the classifier being trained using a data set generated using a policy mapping intermediate results to undesired model behavior including a plurality of enumerated prohibited actions; and

providing data characterizing the determination to a consuming application or process, the consuming application or process:

preventing the multimodal prompt from being ingested by the GenAI model in response to determining that the prompt elicits undesired actions by the GenAI model; and

allowing the multimodal prompt to be ingested by the GenAI model in response to determining that the prompt does not elicit undesired actions by the GenAI model.

Assignments (1)

Continuity (2)

References Cited (204)