Development teams are currently creating dependencies that do not appear in classic software inventories: AI models, tools based on the Model Context Protocol (MCP, an interface through which language models access external tools and data sources), and autonomous agents that work through tasks independently. According to a recent account, these building blocks frequently reach production operation without the security teams knowing about them.

This is precisely where the AI Bill of Materials comes in, or AI-BOM for short. The term is modelled on the software bill of materials: instead of recording only libraries and versions, an AI-BOM documents the inventory of AI components across all clouds in use. Each component is assigned a responsible person, and the inventory can be evidenced for audits.

Why an inventory is the first step

Transparency is the prerequisite for any further control. Anyone who does not know which model answers in which application, which MCP tool accesses which data source and which agent performs tasks in the background can neither assess risks nor shut things down quickly in an emergency.

In practice, an AI-BOM primarily answers three questions:

  • What is running? Which models, tools and agents are in use – including the instances that have arisen in side projects or test environments.
  • Who is responsible? A named responsible person per component prevents responsibilities from becoming orphaned after the project ends.
  • How do we evidence it? The documented inventory serves as evidence for internal audit and external reviews.

For companies with a distributed cloud landscape, the point "across all clouds" is the decisive one. AI functions rarely arise in one place: a model runs with provider A, a vector store with provider B, the agent on its own instance. Without a cross-cutting inventory, every partial picture remains incomplete.

Access rights: platforms are following suit

In parallel with inventorying, the second control level is shifting: the granting of rights on the end device. Apple is tightening the controls for full access to the Mac's hard drive in order to increase security against AI agents. The background is the observation that many users confirm access requests too readily, even though the risks are growing.

This is more than a detail change in an operating system. An agent with full file system access can read, combine and pass on content that was never intended for it. When platform operators raise the hurdles for such approvals, this changes the assumptions on which locally running AI tools are built.

For project managers this means: permissions are not a one-off installation step but an operational matter. What is granted today with a click may be requested again and in more differentiated form after a platform update – with implications for rollout plans and support effort.

Autonomy without clear boundaries

How independently agents can act is shown by an episode surrounding the Openclaw tool. A user reported on X that two Openclaw agents coordinated with each other during his longer absence to find out whether something had happened to him. For six hours the tools occupied themselves with this question – the user was merely asleep.

The story seems curious but illustrates a real mechanism. Agents pursue goals, interpret context and coordinate with one another. If there is no clear definition of what they may do, how long they remain active and when they must ask for clarification, activities arise without a mandate. In this case it remained wasted computing time. In an environment with write permissions, API access or customer data, the same behaviour would be considerably more unpleasant.

The common thread of the three developments

Inventory, granting of rights and agent behaviour are three perspectives on the same gap: AI components spread faster than organisations can record and limit them. The AI-BOM addresses visibility, the tightened platform controls the access level, the Openclaw incident the behaviour level.

What matters is the classification: the sources describe different levels of maturity. The AI-BOM is an organisational concept that a company must introduce itself. The Apple change is a platform measure to which teams react. The Openclaw case is an individual user report and not a systematic finding – it serves as an illustration, not as evidence of a frequency.

What teams can tackle concretely

  • Take stock: Record all models, MCP tools and agents in use, including those that have arisen in pilot projects.
  • Assign responsibility: A named person per component who decides on updates, deactivation and access rights.
  • Minimise rights: Give agents only the access they need for the specific task – treat full access to file systems as an exception, not the standard.
  • Limit runtimes: Define clear abort criteria and time windows so that agents do not remain active for hours unnoticed.
  • Establish evidencing capability: Keep the inventory in such a way that it can be presented at an audit without additional research.

Classification for web and software projects

Anyone integrating AI functions into client projects should maintain the inventory of AI components from the outset and not catch up only at the first audit – the effort is considerably lower at the start of the project.

Permission assignment and agent autonomy belong in the architecture decision, not in the operational phase: which data sources an agent may see and when it stops can be defined cleanly at the design stage, but can only be corrected with difficulty in a running system.For offers and maintenance contracts, this means treating governance effort as a separate line item instead of booking it under "integration".

Sources