Cybersecurity compliance assessment

1 . A computer-implemented method for cybersecurity compliance assessment, the method comprising:

obtaining one or more documents containing a description of an (Information Technology (IT) infrastructure, or of an organization or project relying on an IT infrastructure;

obtaining a plurality of requirements, including a set of cybersecurity requirements, each requirement including a text, the text of each obtained requirement containing one or more sentences;

determining, for each document of the obtained one or more documents, one or more respective document vectors, each respective document vector being a numerical vector representing a chunk of the document, thereby obtaining a plurality of document vectors each representing a respective chunk of a respective document, the determining of each respective document vector comprising:

determining a respective text representation of the respective chunk, the respective text representation of each chunk containing one or more sentences; and

vectorizing the respective text representation of the respective chunk, the vectorizing comprising providing as input the respective text representation of the respective chunk to an embedding module, thereby obtaining as output the respective document vector, wherein the embedding module comprises a first embedding sub-module and a second embedding sub-module, and the embedding module is configured for taking a respective text containing one or more respective sentences as input and for:

by the first embedding sub-module, taking as input each sentence of the one or more respective sentences and determining an embedding vector representing the sentence, thereby obtaining a set of one or more embedding vectors, and

by the second embedding sub-module, taking as input the set of one or more embedding vectors and combining the set of one or more embedding vectors into a single vector representing the respective text provided as input to the embedding module;

determining, for each requirement of the obtained plurality of requirements, a respective requirement vector, the respective requirement vector being a numerical vector representing the text of the requirement, the determining of the respective requirement vector comprising providing as input the text of the requirement to the embedding module, thereby obtaining as output the respective requirement vector;

selecting, for each requirement, one or more respective chunks each of a respective document, the selecting being based on a relevance of the document vectors with respect to the respective requirement vector;

assessing, for each requirement of the obtained plurality of requirements, a level of compliance of the description with the requirement, the assessing being based on the requirement and on the selected one or more respective chunks, the assessing being performed by an assessing module, the assessing module comprising a plurality of Large Language Models (LLMs) configured to cooperate for computing the level of compliance; and

generating digital data that represents a result of the assessment.

2 . The method of claim 1 , wherein at least one chunk of at least one document contains at least one piece of non-textual data, and determining a respective text representation of the respective chunk comprises converting each piece of non-textual data of each chunk into a piece of textual data and including the piece of textual data in the respective text representation of the respective chunk, the converting comprising providing as input each piece of non-textual data to a converting module including a respective LLM.

3 . The method of claim 1 , wherein the first embedding sub-module comprises a trained embedding model with a transformer-based architecture, the trained embedding model having been trained on a training dataset comprising entries, each entry comprising a cybersecurity text.

4 . The method of claim 1 , wherein the selecting of the one or more respective chunks comprises, for each requirement:

computing a similarity score between each document vector and the respective requirement vector;

identifying a number of document vectors having a smallest distance with the respective requirement vector; and

determining among the chunks corresponding to the identified number of document vectors the one or more respective chunks having a highest relevance with respect to the requirement.

5 . The method of claim 4 , wherein the determining among the chunks corresponding to the identified number of document vectors the one or more respective chunks is performed with a respective LLM, the respective LLM having been fine-tuned using cybersecurity text.

6 . The method of claim 1 , wherein the assessing module is configured for performing one or more executions of a block including:

a first LLM which provides as output a respective level of compliance among a set of values including “fully compliant”, “partially compliant”, and “non-compliant”, and

a second LLM configured to be executed when, and only when, the first LLM provides as output “partially compliant”, the second LLM being configured to refine the respective level of compliance;

the level of compliance of the description with the requirement provided by the assessing module being based on the respective level of compliance outputted by each execution of the block.

7 . The method of claim 6 , wherein:

the assessing module is configured for further executing, after each respective execution of the block a third LLM, the third LLM being configured for taking as input the requirement, the selected one or more respective chunks, and the respective level of compliance outputted by the respective execution of the block, and providing as output a respective score of confidence indicating an exactness of the respective level of compliance outputted by the respective execution of the block;

the assessing module is configured for sequentially iterating executions of the block and of the third LLM until the respective score of confidence is higher than a predetermined threshold; and

the level of compliance of the description with the requirement provided by the assessing module is based on the respective level of compliance outputted by a last execution of the block.

8 . The method of claim 6 , wherein:

the assessing module is configured for performing several executions of the block independently and in parallel; and

the level of compliance of the description with the requirement provided by the assessing module is a most common level of compliance among the respective level of compliance outputted by each execution of the block.

9 . The method of claim 1 , wherein the digital data comprises a document data file.

10 . The method of claim 1 , wherein the method further comprises establishing a justification of the result of the assessment with a respective LLM.

11 . The method of claim 1 , wherein the digital data comprises, for at least one requirement, an instruction to increase the level of compliance.

12 . The method of claim 11 , wherein the method comprises establishing the instruction to increase the level of compliance with a respective LLM.

13 . The method of claim 1 , wherein the IT infrastructure is a real-world IT infrastructure, and the digital data comprise, for at least one requirement, configuration data to be deployed in the IT infrastructure to increase the level of compliance.

14 . The method of claim 1 , wherein the digital data comprises a graphical representation of the result, the method further comprises displaying the graphical representation on a display.

15 . The method of claim 1 , wherein the method further comprises creating a data management component comprising at least:

a requirement dataset comprising entries, each entry being indexed by a requirement, and comprising the text of the requirement;

a description dataset comprising entries, each entry being indexed by a chunk of one of the documents, and comprising the determined text representation of the chunk; and

a vector dataset comprising requirement entries and description entries, each requirement entry comprising:

a pointer to an entry of the requirement dataset; and

a respective requirement vector having been determined from the requirement indexing the pointed entry;

each description entry comprising:

a pointer to an entry of the description dataset; and

a respective description vector having been determined from the description indexing the pointed entry.

16 . A non-transitory computer readable medium having recorded thereon instructions configured for, when executed by one or more processors, causing performance of a method for cybersecurity compliance assessment, the method comprising:

obtaining one or more documents containing a description of an Information Technology (IT) infrastructure, or of an organization or project relying on an IT infrastructure;

obtaining a plurality of requirements, including a set of cybersecurity requirements, each requirement including a text, the text of each obtained requirement containing one or more sentences;

determining, for each document of the obtained one or more documents, one or more respective document vectors, each respective document vector being a numerical vector representing a chunk of the document, thereby obtaining a plurality of document vectors each representing a respective chunk of a respective document, the determining of each respective document vector comprising:

determining a respective text representation of the respective chunk, the respective text representation of each chunk containing one or more sentences; and

vectorizing the respective text representation of the respective chunk, the vectorizing comprising providing as input the respective text representation of the respective chunk to an embedding module, thereby obtaining as output the respective document vector, wherein the embedding module comprises a first embedding sub-module and a second embedding sub-module, and the embedding module is configured for taking a respective text containing one or more respective sentences as input and for:

by the first embedding sub-module, taking as input each sentence of the one or more respective sentences and determining an embedding vector representing the sentence, thereby obtaining a set of one or more embedding vectors, and

by the second embedding sub-module, taking as input the set of one or more embedding vectors and combining the set of one or more embedding vectors into a single vector representing the respective text provided as input to the embedding module;

determining, for each requirement of the obtained plurality of requirements, a respective requirement vector, the respective requirement vector being a numerical vector representing the text of the requirement, the determining of the respective requirement vector comprising providing as input the text of the requirement to the embedding module, thereby obtaining as output the respective requirement vector;

selecting, for each requirement, one or more respective chunks each of a respective document, the selecting being based on a relevance of the document vectors with respect to the respective requirement vector;

assessing, for each requirement of the obtained plurality of requirements, a level of compliance of the description with the requirement, the assessing being based on the requirement and on the selected one or more respective chunks, the assessing being performed by an assessing module, the assessing module comprising a plurality of Large Language Models (LLMs) configured to cooperate for computing the level of compliance; and

generating digital data that represents a result of the assessment.

17 . A computer system comprising one or more processors coupled to memory, the memory having recorded thereon instructions configured for, when executed by one or more processors, causing performance of a method for cybersecurity compliance assessment, the method comprising:

obtaining one or more documents containing a description of an Information Technology (IT) infrastructure, or of an organization or project relying on an IT infrastructure;

obtaining a plurality of requirements, including a set of cybersecurity requirements, each requirement including a text, the text of each obtained requirement containing one or more sentences;

determining, for each document of the obtained one or more documents, one or more respective document vectors, each respective document vector being a numerical vector representing a chunk of the document, thereby obtaining a plurality of document vectors each representing a respective chunk of a respective document, the determining of each respective document vector comprising:

determining a respective text representation of the respective chunk, the respective text representation of each chunk containing one or more sentences; and

vectorizing the respective text representation of the respective chunk, the vectorizing comprising providing as input the respective text representation of the respective chunk to an embedding module, thereby obtaining as output the respective document vector, wherein the embedding module comprises a first embedding sub-module and a second embedding sub-module, and the embedding module is configured for taking a respective text containing one or more respective sentences as input and for:

by the first embedding sub-module, taking as input each sentence of the one or more respective sentences and determining an embedding vector representing the sentence, thereby obtaining a set of one or more embedding vectors, and

by the second embedding sub-module, taking as input the set of one or more embedding vectors and combining the set of one or more embedding vectors into a single vector representing the respective text provided as input to the embedding module;

determining, for each requirement of the obtained plurality of requirements, a respective requirement vector, the respective requirement vector being a numerical vector representing the text of the requirement, the determining of the respective requirement vector comprising providing as input the text of the requirement to the embedding module, thereby obtaining as output the respective requirement vector;

selecting, for each requirement, one or more respective chunks each of a respective document, the selecting being based on a relevance of the document vectors with respect to the respective requirement vector;

assessing, for each requirement of the obtained plurality of requirements, a level of compliance of the description with the requirement, the assessing being based on the requirement and on the selected one or more respective chunks, the assessing being performed by an assessing module, the assessing module comprising a plurality of Large Language Models (LLMs) configured to cooperate for computing the level of compliance; and

generating digital data that represents a result of the assessment.