# US 2025/0165617 A1 · Generative cybersecurity exploit discovery and evaluation

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

- Number: US 2025/0165617 A1
- Document ID: US20250165617A1
- Status: Published application
- Kind: A1
- Application number: 18/900,216
- Filed: 2024-09-27
- Published: 2025-05-22
- Earliest claimed: 2023-01-19
- Assignees: Citibank, N.A.
- Applicants: Citibank, N.A.
- Inventors: William Franklin Cameron; Pramod Goyal; Prithvi Narayana Rao; Manjit Rajaretnam; Miriam Silver
- Web page: https://patentgrep.com/patent/US20250165617A1

## CPC

- G06F 21/577: {Assessing vulnerabilities and evaluating computer system security}
- G06F 21/552: {involving long-term monitoring or reporting}
- G06F 2221/034: Test or assess a computer or a system
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Described herein are systems and methods for discovering and proactively mitigating previously unknown security vulnerabilities. The systems and methods herein can utilize security vulnerability information to discover potential security threats and can utilize this information to generate an attack using a machine learning model, such as a large language model. Generated attacks can be carried out to assess impact of a security vulnerability. An output can be provided that represents the assessed impact. In some implementations, the systems and methods herein generate patches or other mitigations for security vulnerabilities, which can be tested and deployed to address security vulnerabilities.

## Claims

### Claim 1

A method for automatic vulnerability discovery and mitigation, the method comprising:
determining an application of interest for vulnerability detection and mitigation,
wherein the application of interest is determined based on at least one of:
receiving a user indication, identifying a newly installed application, identifying an application update, or identifying an application configuration change;
determining application analysis data of the application, wherein the application analysis data comprises one or more of: static analysis data, source code analysis data, dynamic analysis data, or network activity data;
generating, based on the application analysis data, a set of data comprising a representation of at least a portion of the application analysis data;
applying a trained machine learning model on the set of data to generate a set of outputs indicative of a potential vulnerability;
determining, based at least in part on the set of outputs, that the potential vulnerability is a vulnerability, wherein determining that the potential vulnerability is a vulnerability comprises:
determining an exploit for the potential vulnerability, wherein determining the exploit for the potential vulnerability comprises at least one or identifying an existing exploit based at least in part on at least one of the set of data or set of outputs or generating an exploit based at least in part on at least one of the set of data or the set of outputs;
executing the exploit; and
analyzing a result of the exploit to determine that the exploit was successful;
determining, for the vulnerability, a mitigation action, wherein the mitigation action is determined at least in part based on providing the set of outputs to a second machine learning model;
applying the mitigation action.

### Claim 2

A method for automatic vulnerability discovery and mitigation, the method comprising:
determining an application of interest for vulnerability detection and mitigation,
wherein the application of interest is determined based on at least one of:
receiving a user indication, identifying a newly installed application, identifying an application update, or identifying an application configuration change;
determining application analysis data of the application, wherein the application analysis data comprises one or more of: static analysis data, source code analysis data, dynamic analysis data, or network activity data;
generating, based on the application analysis data, a set of data comprising a representation of at least a portion of the application analysis data;
applying a trained machine learning model on the set of data to generate a set of outputs indicative of a potential vulnerability;
determining, based at least in part on the set of outputs, that the potential vulnerability is a vulnerability.

### Claim 3 (depends on 2)

The method of claim 2, wherein the trained machine learning model is configured to generate outputs indicative of at least one of: a type of vulnerability, an attack vector, required privileges, or user interaction requirements.

### Claim 4 (depends on 2)

The method of claim 2, wherein determining that the potential vulnerability is a vulnerability comprises:
generating, based at least in part on the set of outputs, a software attack comprising executable code,
wherein the executable code is generated using at least one large language model,
wherein the at least one large language model is configured to:
access supporting information from a data store, the supporting information comprising a code snippet used for generating the executable code; and
receive a prompt comprising at least an indication of the set of outputs;
executing the executable code against a potentially vulnerable system; and
determining based on a result of executing the executable code, that the potentially vulnerable system is vulnerable.

### Claim 5 (depends on 2)

The method of claim 2, further comprising:
determining, for the vulnerability, a mitigation action, wherein the mitigation action is determined at least in part based on providing the set of outputs to a second machine learning model.

### Claim 6 (depends on 2)

The method of claim 2, further comprising:
determining an assessed risk level for the potential vulnerability.

### Claim 7 (depends on 2)

The method of claim 2, wherein determining application analysis data of the application comprises performing static analysis, wherein performing the static analysis comprises:
disassembling the application using a disassembler;
identifying one or more strings in the application; and
identifying one or more execution paths of the application.

### Claim 8 (depends on 2)

The method of claim 2, wherein determining application analysis data of the application comprises performing dynamic analysis, wherein performing the dynamic analysis comprises:
executing the application; and
monitoring, during the executing, operations of the application, the operations comprising one or more of: a file operation, a network operation, or a memory operation.

### Claim 9 (depends on 2)

The method of claim 2, wherein the trained machine learning model comprises a large language model, wherein the large language model is configured to discover vulnerabilities using one of: static analysis data, dynamic analysis data, code analysis data, or network analysis data.

### Claim 10 (depends on 2)

The method of claim 2, wherein the trained machine learning model comprises a plurality of large language models, wherein each large language model of the plurality of large language models is configured to perform a different vulnerability discovery task of a set of vulnerability discovery tasks.

### Claim 11 (depends on 10)

The method of claim 10, wherein the set of vulnerability discovery tasks includes vulnerability discovery in different assessment domains.

### Claim 12 (depends on 10)

The method of claim 10, wherein the set of vulnerability discovery tasks includes vulnerability discovery using different types of vulnerability discovery data, wherein the different types of vulnerability discovery data comprise at least one of static analysis data, dynamic analysis data, source code analysis data, or network analysis data.

### Claim 13 (depends on 2)

The method of claim 2, further comprising:
determining a platform associated with the application, wherein the platform indicates an ecosystem of computing resources associated with the application;
identifying a set of assessment domains associated with the platform by comparing a platform identifier of the platform to an assessment domain mapping data structure;
generating an attack based on the vulnerability using a large language model;
executing the attack against the platform;
determining, based on a result of executing the attack, an assessment domain impact level for an assessment domain of the set of assessment domains associated with the platform; and
generating for output, a representation indicating at least one assessment domain-specific impact level.

### Claim 14

A system for automatic vulnerability discovery and mitigation, the system comprising:
at least one hardware processor; and
at least one non-transitory memory storing instructions which, when executed by the at least one hardware processor, cause the system to:
determine an application of interest for vulnerability detection and mitigation,
wherein the application of interest is determined based on at least one of: receiving a user indication, identifying a newly installed application, identifying an application update, or identifying an application configuration change;
determine application analysis data of the application, wherein the application analysis data comprises one or more of: static analysis data, source code analysis data, dynamic analysis data, or network activity data;
generate, based on the application analysis data, a set of data comprising a representation of at least a portion of the application data;
apply a trained machine learning model on the set of data to generate a set of outputs indicative of a potential vulnerability; and
determine, based at least in part on the set of outputs, that the potential vulnerability is a vulnerability.

### Claim 15 (depends on 14)

The system of claim 14, wherein determining that the potential vulnerability is a vulnerability comprises:
generating, based at least in part on the set of outputs, a software attack comprising executable code,
wherein the executable code is generated using at least one large language model,
wherein the at least one large language model is configured to:
access supporting information from a data store, the supporting information comprising a code snippet used for generating the executable code; and
receive a prompt comprising at least an indication of the set of outputs;
executing the executable code against a potentially vulnerable system; and
determining based on a result of executing the executable code, that the potentially vulnerable system is vulnerable.

### Claim 16 (depends on 14)

The system of claim 14, wherein the instructions are further configured to cause the system to:
determine, for the vulnerability, a mitigation action, wherein the mitigation action is determined at least in part based on providing the set of outputs to a second machine learning model.

### Claim 17 (depends on 14)

The system of claim 14, wherein the trained machine learning model comprises a large language model, wherein the large language model is configured to discover vulnerabilities using one of: static analysis data, dynamic analysis data, code analysis data, or network analysis data.

### Claim 18 (depends on 14)

The system of claim 14, wherein the trained machine learning model comprises a plurality of large language models, wherein each large language model of the plurality of large language models is configured to perform a different vulnerability discovery task of a set of vulnerability discovery tasks.

### Claim 19 (depends on 18)

The system of claim 18, wherein the set of vulnerability discovery tasks includes vulnerability discovery using different types of vulnerability discovery data, wherein the different types of vulnerability discovery data comprise at least one of static analysis data, dynamic analysis data, source code analysis data, or network analysis data.

### Claim 20 (depends on 14)

The system of claim 14, wherein the instructions are further configured to cause the system to:
determine a platform associated with the application, wherein the platform indicates an ecosystem of computing resources associated with the application;
identify a set of assessment domains associated with the platform by comparing a platform identifier of the platform to an assessment domain mapping data structure;
generate an attack based on the vulnerability using a large language model;
execute the attack against the platform;
determine, based on a result of executing the attack, an assessment domain impact level for an assessment domain of the set of assessment domains associated with the platform; and
generate for output, a representation indicating at least one assessment domain-specific impact level.
