LLM Supply Chain Attestations for Secure Model Provenance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The potential for misuse and errors in large language models (LLMs) poses significant security and privacy concerns, including the creation of deceptive and inaccurate content, perpetuation of biases, and leakage of sensitive information, necessitating a need for secure and safe programmatic usage throughout the development process.

Innovation Solution

Extending the software bill of materials (SBOM) with LLM-specific provenance parameters and signed attestations, attaching an LLM-extended SBOM to applications, and performing LLM-specific tests at various development stages to ensure security and safety.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If LLMs are used to generate text content, then utility and coherence are improved, but security risks and potential misuse increase

Engineering Contradiction:
Improvetext generation capabilityVSAvoidsecurity risks
Core Design Contradiction:
Ease of operationVSObject-affected harmful factors

Solution Approach 1:

The patent introduces an intermediary verification system that sits between the LLM and the end user. This system includes a digital signature verification module that checks attestations about the LLM's behavior, safety, and provenance before allowing text generation output to be displayed or used. The intermediary validates that the LLM has not engaged in harmful behaviors such as generating deceptive content, leaking sensitive information, or producing biased output.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements preliminary actions by performing security verification and validation before the LLM generates text content. The system digitally signs attestations regarding the LLM's training data provenance, model architecture integrity, and safety performance metrics in advance. These pre-verified attestations are then attached to the LLM output, allowing downstream systems to verify safety properties before processing or displaying generated content.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If comprehensive LLM testing is performed at various development stages, then security and safety are improved, but development time and complexity increase

Engineering Contradiction:
Improvesecurity verificationVSAvoidtesting infrastructure
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms the verification process by changing parameters from manual security auditing to automated digital signature verification. Instead of complex human review processes, the system uses cryptographic signatures and machine-readable attestations that can be automatically verified by software. This parameter change maintains high security verification while reducing the operational complexity of the testing infrastructure.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates simplified copies or representations of complex security verification processes through digital attestations. Rather than requiring full replication of testing environments and procedures, the system generates compact digital certificates that encapsulate verification results. These attestations serve as verified copies of security properties that can be easily transmitted and checked without reproducing the entire testing infrastructure.

Inventive Principle:
Principle #26Copying

3Loss of information

If LLM provenance tracking is implemented, then data security and accountability are improved, but system complexity and overhead increase

Engineering Contradiction:
Improvedata provenance trackingVSAvoidprovenance management system
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent extracts provenance information from the complex LLM development process and separates it into distinct, verifiable components. The system pulls out key provenance elements such as training data sources, model architecture details, and safety validation results, then digitally signs each component independently. This extraction approach maintains comprehensive provenance tracking while reducing system complexity by modularizing the tracking infrastructure into discrete, manageable units.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentUS12579256B2Large language model (LLM) supply chain security
Publication Date: 2026.03.17 AMERICAN EXPRESS TRAVEL RELATED SERVICES CO INC
  • US12579256B2 patent drawing
  • US12579256B2 patent drawing
  • US12579256B2 patent drawing

AI summary

Disclosed are various approaches for large language model (LLM) supply chain security. In one example, an LLM-extended software bill of materials can be extended to provide LLM specific supply chain information for an LLM application that communicates with an LLM service. The LLM-extended software bill of materials can be attached to the LLM application. An LLM specific security test can be performed on the LLM application. A signed LLM security test attestation can be attached to the LLM-extended software bill of materials based on completion of the automated LLM security test. The LLM application or the LLM-extended software bill of materials can be published or transmitted to a predetermined network endpoint.