# A case for how to shape ‘ingredient lists’ for AI models > Source: > Published: 2026-06-16 16:00:00+00:00 # A case for how to shape ‘ingredient lists’ for AI models A policy paper published Tuesday advocates for software bills of materials (SBOMs) for artificial intelligence as a mechanism for reducing cyber risk and improving transparency, and seeks to give lawmakers, federal agencies and others a roadmap on how to proceed. The SBOM, commonly described as an inventory of software ingredients, emerged in the 2010s and has expanded beyond software to include hardware and [AI](https://cyberscoop.com/tag/ai/). But [the paper](https://securityandtechnology.org/virtual-library/policy-memo/driving-ai-transparency/) from the [Institute for Security and Technology](https://cyberscoop.com/tag/institute-for-security-and-technology/), which CyberScoop is the first to report on, argues that AIBOMS require foundational work before they can be widely implemented. This comes as some companies are already offering AIBOM services and [other organizations](https://cyberscoop.com/g7-cisa-ai-sbom-security-guidance/) are actively shaping AIBOM policy. “What we’re worried about is we would end up in a ‘fire, ready, aim’ situation where everyone was doing it, but we were all doing slightly different things,” said a co-author of the paper, Allan Friedman, who has worked on SBOMs in multiple U.S. government roles. “If we don’t have a shared vision, it becomes a lot harder to have a coherent policy. It becomes a lot harder to have common tools and interoperable data and it becomes a lot harder to use the data that we’re tracking to actually deliver on the promise of supply chain transparency.” The idea for the paper sprung from discussions with Hill aides and Pentagon staffers, Friedman said, and people like them are the target audience as well. A key premise is that AIBOM policy needs to explore the topic from two sides. “How do you solve the chicken-and-egg issue, where no one’s providing the data, so no one’s asking for it, and no one’s asking for it, so no one’s providing it?” Friedman told CyberScoop. “The answer is, you have to go from both supply and demand.” On the supply side, “An AIBOM should capture relevant details about the models and datasets used for training, fine-tuning, evaluation, validation, testing, retrieval, grounding, augmentation, or other model development or operational purposes,” the paper suggests. “The demand side begins with some form of forcing function or requirement that organizations understand what is in the products they manufacture and sell,” it states, with one such requirement potentially being an industry mandate to require the tracking of system components — for example, like the “lightweight” standards used in the payment card industry on data security that isn’t overly exact about how components should be tracked. But it could also include government regulations or contracting conditions, Friedman argues with his Institute for Security and Technology colleague Nick Leiserson. (The scope of government directives on AI is a topic of considerable debate [on Capitol Hill](https://www.politico.com/news/2026/06/13/republicans-ai-josh-hawley-tech-republicans-artificial-intelligence-00961315) and [within the Trump administration](https://cyberscoop.com/donald-trump-white-house-ai-executive-order-scaled-back/) right now.) Friedman said the paper isn’t meant to be the be-all, end-all, and acknowledged the prior work of organizations like the Open Worldwide Application Security Project ([OWASP](https://cyberscoop.com/tag/owasp/)) and Linux Foundation. “We’re not saying this is a brand new topic, nor are we saying that AIBOM will solve all AI security issues,” he said. “I’ve been fighting this fight for [SBOM](https://cyberscoop.com/tag/sbom/) for a decade. You know, SBOM will not pick up your dry cleaning.” And as AI continues to evolve rapidly, that means papers like the one published Tuesday are just at the beginning of the discussion, Friedman said.