SIAgentNews.com
← All news

New AI Bill of Materials helps track hidden components in software

Extended software inventories are being developed to track AI models and datasets, helping companies meet upcoming security reporting requirements.

· First reported by heise online

24 hoursTimeframe to report vulnerabilities

The short version

  • Standard SBOMs typically list packages and libraries but omit AI models, weights, and training datasets.
  • The EU Cyber Resilience Act will require manufacturers to report exploited vulnerabilities within 24 hours starting in 2026.
  • New standards from CycloneDX and SPDX allow for specific AI and dataset profiles to be included in inventories.
  • Tools like Snyk and cdxgen are being used to identify AI agents and tools within source code.
  • A complete AI inventory must also account for development environment extensions and 'shadow AI' in production.

Why it matters

As AI becomes a standard part of software, companies need a way to know exactly what models and data they are using. Without an AI-BOM, it is nearly impossible to quickly identify which products are affected by a specific security flaw in a model.

The visibility gap in software

Software Bill of Materials (SBOM) is a standard list of every package and library used to build a piece of software. It acts like an ingredient list, allowing developers to see what is inside their applications to ensure security and compliance. However, as companies integrate artificial intelligence, these traditional lists are failing to capture the full picture.

In a recent example, a development team built a customer service assistant using a model from Hugging Face. The system connected several servers for ticket searching and calendar access, and even utilized a code block generated by an AI assistant. While the final software registry listed 812 packages, it completely missed the AI model, its weights, the training data, and the various AI extensions.

This oversight creates a significant blind spot for security teams. Because these AI components do not appear in standard dependency files, they remain invisible to traditional inventory tools. This makes it difficult for organizations to manage the risks associated with the specific AI building blocks they choose to deploy.

Meeting new legal requirements

The push for better AI tracking is being driven by upcoming regulations like the Cyber Resilience Act (CRA). This law will require manufacturers to report exploited vulnerabilities within 24 hours of discovery. This mandate applies even to older products already in use by customers, creating a high-pressure environment for software providers.

If a company falls under these reporting rules, their first priority will be identifying which components are inside which products. Because AI components are not exempt from these rules, companies must be able to pinpoint them quickly. Without a clear inventory, a company might struggle to determine if a newly discovered flaw in a specific model affects their software.

As heise online reported, the goal is to turn these documents into functional tools. Many organizations currently generate SBOMs just to satisfy compliance checklists without actually analyzing the data. To be useful, these inventories must be automatically cross-referenced with vulnerability databases to identify affected products in minutes.

Standards for AI transparency

To bridge the gap between standard software lists and AI needs, industry groups are creating new formats. CycloneDX is introducing a way to describe models using a Model Card and a proof of origin. Meanwhile, SPDX 3.0.1 is adding specific profiles for both AI models and the datasets used to train them.

These standards are becoming important for regulatory compliance. For instance, the BSI guideline TR-03183-2 specifically requires the use of either CycloneDX 1.6 or SPDX 3.0.1. By adopting these formats, developers can ensure that AI-specific data is structured in a way that security tools can actually read and process.

The transition from a standard SBOM to an AI-BOM involves several technical steps. It requires generating the data during the build process, securing the origin of the components with digital signatures, and enforcing these rules within the development pipeline. This ensures that every AI component is accounted for before the software is ever released to the public.

Tools for identifying AI components

Several tools are emerging to help developers automate the discovery of AI components within their code. The Snyk tool includes an aibom command that can find models, agents, and Model Context Protocol (MCP) tools. Additionally, a tool called cdxgen can read model files to provide a checksum, which helps verify the integrity of the AI components.

Security is further bolstered by the OpenSSF Model Signing project. This initiative binds model weights to a specific identity, making it harder for malicious components to be swapped into a system. These tools allow developers to move beyond manual tracking and toward an automated system of verification for their AI assets.

However, a complete inventory requires looking beyond just the source code. Developers must also account for AI extensions used in their development environments. Furthermore, organizations need to actively search for 'shadow AI'—AI tools or models used by employees that have not been officially approved or vetted by the company's security team.

The path to automated security

The ultimate goal of the AI-BOM is to create a system where security is proactive rather than reactive. By integrating the inventory into the CI/CD pipeline, scripts can automatically decide whether to approve or block a build based on the components it contains. This prevents unvetted or risky AI models from ever reaching the production environment.

For the reader, this means that the software they use will eventually have a much higher level of transparency. While it adds a layer of complexity for developers today, it builds a foundation for safer AI integration. It ensures that when a vulnerability is found, the response can be immediate and precise rather than a guessing game.

Moving forward, the focus will shift from simply creating these documents to effectively using them. The real value lies in the ability to query a database and see exactly which products are at risk. This shift will define how companies manage the lifecycle of AI-powered software in a regulated world.

Sources and links

First reported by heise online. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news