Neural dialogue system for security posture management
A neural dialogue system has been designed to present a conversational user interface for managing security posture of an organization. The neural dialogue system determines intent and extracts entity names from a user input. The neural dialogue system grounds the entity names to an organizational context and maps the intent to a defined functionality related to security posture management for intent realization. Some of these defined functions involve editing the ordered ruleset. When realization of an intent involves editing the ordered ruleset, the neural dialogue system determines various ruleset command sequences that implement the ruleset editing and corresponding impacts on the ordered ruleset (e.g., creation of ruleset anomalies). The dialogue system collects the information retrieved based on the intent and grounded entities, including any ruleset impact assessment, and generates a response.
1 . A method comprising:
using a first language model to determine from a user input a first intent and a first set of one or more entity identifiers;
grounding the first set of entity identifiers to a context of an organization to generate a second set of one or more entity identifiers;
determining first information for responding to the user input based, at least in part, on at least one of querying metadata of a security ruleset of the organization and analyzing the security ruleset, wherein the querying and/or analyzing is based, at least in part, on the first intent and the second set of entity identifiers,
wherein determining the first information comprises,
determining whether structural metadata of the security ruleset would satisfy the first intent;
based on determining that structural metadata of the security ruleset would satisfy the first intent, invoking a function with at least one of the second set of entity identifiers as an argument to query the structural metadata and obtain at least a part of the first information; and
based on determining that structural metadata of the security ruleset would not satisfy the first intent, determining a set of one or more commands to perform on the security ruleset that satisfies the first intent and yields at least a part of the first information; and
communicating the first information for responding to the user input.
2 . The method of claim 1 , wherein determining a set of one or more commands to perform on the security ruleset comprises:
selecting multiple of a plurality of ruleset commands;
for each selection,
determining a sequence of ruleset commands starting with the selected ruleset command and based on the plurality of ruleset commands;
assessing impact of the sequence of ruleset commands on the security ruleset; and
determining which of the sequences of ruleset commands has a lowest impact, wherein the set of one or more commands is the sequence of ruleset commands with the lowest impact.
3 . The method of claim 2 , wherein assessing impact of each sequence of ruleset commands comprises, for each sequence of ruleset commands, determining state of the security ruleset that would result if the sequence of ruleset commands is performed and corresponding impact after performance of each ruleset command in the sequence of ruleset commands.
4 . The method of claim 2 , wherein assessing impact of each sequence of ruleset commands comprises determining ruleset anomalies that would be created.
5 . The method of claim 2 , wherein communicating the first information for responding to the user input comprises indicating a recommendation based on the set of one or more commands.
6 . The method of claim 5 , wherein communicating the first information comprises indicating the impact assessment for the set of commands.
7 . The method of claim 1 , wherein grounding the first set of entity identifiers to a context of the organization to generate a second set of one or more entity identifiers comprises, for each of the first set of entity identifiers, generating a representation based on natural language processing and determining a most similar of a plurality of entity identifier representations previously generated from context data of the organization.
8 . The method of claim 7 , wherein determining the most similar of the plurality of entity identifier representations comprises determining similarity according to priority of the context data.
9 . The method of claim 1 , wherein the context of the organization further comprises at least one of configuration management data and content management data.
10 . A non-transitory, machine-readable medium having program code stored thereon, the program code comprising instructions to:
invoke a first language model on a user input to determine a first intent and a first set of one or more entity identifiers for the user input;
ground the first set of entity identifiers to a context of an organization to generate a second set of one or more entity identifiers;
determine whether structural metadata of a ruleset would satisfy the first intent;
based on a determination that structural metadata of the ruleset would satisfy the first intent, invoke a function with at least one of the second set of entity identifiers as an argument to query the structural metadata;
based on a determination that structural metadata of the ruleset would not satisfy the first intent, determine a set of one or more ruleset commands based, at least in part, on the first intent and the second set of entity identifiers; and
generate a response for the user input based, at least in part, on the set of one or more ruleset commands.
11 . The non-transitory machine-readable medium of claim 10 , wherein the instructions to determine a set of one or more ruleset commands based, at least in part, on the first intent and the second set of entity identifiers comprise instructions to:
select multiple of a plurality of ruleset commands;
for each selection,
determine a sequence of ruleset commands that starts with the selected ruleset command and based on the plurality of ruleset commands;
assess impact of the sequence of ruleset commands on the ruleset; and
determine which of the sequences of ruleset commands has a lowest impact, wherein the set of one or more commands is the sequence of ruleset commands with the lowest impact.
12 . The non-transitory machine-readable medium of claim 11 , wherein the instructions to assess impact of each sequence of ruleset commands on the ruleset comprise instructions to determine state of the ruleset that would result if the sequence of ruleset commands is performed and which, if any, anomaly or anomalies would be created.
13 . The non-transitory machine-readable medium of claim 10 , wherein the instructions to ground the first set of entity identifiers to a context of the organization to generate a second set of one or more entity identifiers comprise instructions to, for each of the first set of entity identifiers, generate a representation based on natural language processing and determine a most similar of a plurality of entity identifier representations previously generated from context data of the organization.
14 . An apparatus comprising:
a processor;
a machine-readable medium having instructions stored thereon, the instructions executable by the processor to cause the apparatus to,
invoke a neural language model to generate a task-based intent of an utterance in a dialogue and identify a set of one or more entities from the utterance;
determine organization-based entity identifiers for the identified set of entities based, at least in part, on data of an organization;
determine whether the task-based intent corresponds to analyzing an ordered ruleset of the organization;
based on a determination that the task-based intent corresponds to analyzing the ordered ruleset,
determine a plurality of sequences of commands to perform on the ordered ruleset based, at least in part, on the task-based intent and the set of entities; and
for at least a subset of the plurality of sequences of commands, determine anomaly-based impact on the ordered ruleset if the sequence of commands is performed; and
generate a response corresponding to the utterance based, at least in part, on the determined anomaly-based impact on the ordered ruleset.
15 . The apparatus of claim 14 , wherein the instructions to determine organization-based entity identifiers for the identified set of entities comprise the instructions being executable by the processor to cause the apparatus to, for each of the identified set of entities, generate a representation based on natural language processing and determine a most similar of a plurality of organization-based entity identifier representations previously generated from context data of the organization.
16 . The apparatus of claim 14 , wherein the instructions to determine anomaly-based impact on the ordered ruleset for each of the subset of the sequences of commands if performed comprise instructions executable by the processor to cause the apparatus to, for each of the subset of the sequences of commands, determine state of the ordered ruleset that would result if the sequence of commands is performed and corresponding anomaly-based impact after performance of each ruleset command in the sequence of ruleset commands.
17 . The apparatus of claim 14 , wherein the instructions to generate the response comprise instructions executable by the processor to cause the apparatus to indicate a recommendation for one of the subset of sequences of commands based on the anomaly-based impacts.
18 . The apparatus of claim 14 , wherein the machine-readable medium further has stored thereon instructions executable by the processor to cause the apparatus to query structural metadata of the ordered ruleset based on a determination that the task-based intent does not correspond to analyzing the ordered ruleset.
19 . The non-transitory machine-readable medium of claim 11 , wherein the instructions to generate a response for the user input comprise instructions to indicate the impact assessment for the set of ruleset commands and to indicate a recommendation based on the set of one or more ruleset commands.
20 . The non-transitory machine-readable medium of claim 10 , wherein the instructions to determine whether structural metadata of the ruleset would satisfy the first intent comprises the instructions to determine whether the first intent is to query the ruleset or to edit or configure the ruleset, wherein structural metadata can satisfy a query type of intent.