AI and Virtual Machine Accountability
A proposed worldwide standard for keeping verifiable records of AI and virtual-machine activity private during routine use and available to authorized investigators when there is a credible incident.
Explore the partner pilotWhat the standard covers
The same rules would apply across vendors and countries, on local and cloud systems. The goal is coverage of every VM, every microchip, every LLM, and every individual model input and output token.
Every microchip
Give chips and firmware verifiable identities and track which versions a workload uses.
Every virtual machine
Record lifecycle events and link observed activity to the workload, actor, action, and policy decision. Surface gaps in collection.
Every LLM
Identify the version of each model sold, distributed, or run, and verify the available provenance before use.
Every individual token
Bind each input and output token to its request, position, and model version through a protected commitment.
Software and data provenance
Track the software, model-building tools, and datasets used to create or run a workload.
How accountability would work
The proposed design would verify evidence before a VM starts, a model loads, or an agent receives permission to act, using authenticated tunnels, isolated runtimes and scoped permissions. Defensive decoys are a proposed future control; none has been built.
Witnessed ledger checkpoints
Signed checkpoints have passed component tests in the lab. Independent witnesses are proposed and not yet demonstrated; with them, changes, missing entries or conflicting views would be detectable.
Private during routine use
Keep prompts, outputs, credentials, and identifiable usage records out of shared ledgers. Limit and log access to evidence during an authorized investigation.
Independent verification
Let a second operator check signed evidence and reconcile expected activity against what was actually recorded.
Partner pilot
BB2G is seeking infrastructure operators, security researchers, and public-interest partners to verify the same VM and model evidence independently, detect missing or altered records, and investigate a policy violation without exposing private activity.