AI fleets need a risk meter before spending what cannot be undone

What the source reports

The arXiv abstract says fleets of AI systems that use tools can cause effects that cannot be fully undone. Examples include moving money, deploying code, deleting data, and disclosing information. Current controls check one effect at a time. So approved tools can still overdraw a principal’s shared risk limit under one trigger. The paper proposes an irreversibility budget. A trusted runtime keeps a residual value-at-risk account for each principal across AI tools, job steps, and tenants. Each effect is charged for residual loss. The marginal effect is denied if the total would exceed the budget. In a controlled study, per-effect gates allowed overdraws up to 48 times the tenant’s risk limit. Correctly charged budget runs stayed within that limit. The open problem is cautious pricing that accounts for dependencies.