Skip to main content

AI-BOM - AI Bill of Materials

Path: /assets/ai-bom (Assets → AI-BOM)

The AI-BOM answers a question most teams cannot: which of our applications use AI, and what do they pull in? It surfaces "shadow AI" - the model SDKs, agent frameworks and vector stores that arrive as ordinary dependencies and never get a security review.

Where the data comes from

The AI-BOM is not a separate scan. It is a lens over the SCA component inventory you have already collected: every dependency parsed from your repos' manifests is classified against a curated catalog of AI packages. Nothing new is fetched, and no source code is read for this view - it works entirely from declared dependencies (name, version, ecosystem, direct/transitive, source file).

What it classifies

Each AI-related dependency is bucketed into one of five categories:

CategoryExamples
Model providerOpenAI, Anthropic, Cohere, Mistral, Google generative-ai / aiplatform, Replicate, Together, Groq, the Vercel AI SDK, tiktoken
Agent frameworkLangChain / langgraph / langsmith, LlamaIndex, Haystack, CrewAI, AutoGen, Semantic Kernel, Instructor, Guidance, DSPy, the MCP SDK
ML runtimeHugging Face transformers / sentence-transformers, PyTorch, TensorFlow, ONNX, Ollama, llama.cpp, vLLM
Vector databaseChroma, Pinecone, Weaviate, Qdrant, FAISS, pgvector, Milvus
GuardrailGuardrails AI, NeMo Guardrails, Ragas, LLM Guard

Deliberately generic names are matched exactly (not by substring) so an unrelated package is never mislabelled as AI.

Reading the AI-BOM

  • Estate view - the apps that use AI, each with its providers, categories and package count, plus a summary: how many apps use AI, how many AI packages in total, and the spread by provider and by category.
  • Per-application view - one app's AI packages grouped by category, with totals and the provider list.
Why it matters

A vector store or a model SDK is a dependency like any other - it can carry a CVE, be reachable or not, and it expands your data-flow surface. The AI-BOM lets you answer "where is AI in our estate, and is any of it risky?" without a spreadsheet.

Related: AI config scan covers the instruction side of AI risk, and slopsquat detection covers AI-suggested packages that should not be trusted.