On October 15, 2026, the Financial Stability Board (FSB) finalized 12 non-binding "sound practices" for how financial firms adopt AI. Coverage of the agentic AI guidance highlights hard limits on the tools and data an agent can reach, records that let a firm rebuild what an agent did, a documented kill switch, and human sign-off for large transactions.
Every one of those controls assumes you can point to one specific agent. Limits, audit trails and kill switches all depend on each agent having its own identity and a named owner.
Last week the Financial Stability Board (FSB) finalized 12 "sound practices" for how banks and other financial firms adopt AI. The guidance is not binding. But regulators tend to treat sound practices as the bar over time.
Read the parts about agentic AI (AI that carries out multi-step tasks on its own) and one theme stands out: know exactly what each agent is allowed to do, and prove what it did.
Hard limits on the tools and data an agent can reach.
Records good enough to rebuild what the agent saw, decided, and did.
A documented kill switch.
Human sign-off for large transactions.
Every one of those starts with identity. You can't limit access for an agent you can't name. You can't rebuild a decision trail if five agents share one API key. And a kill switch only works if you know which credential to cut.
I spent years in identity and certificate management. The lesson carries over: Identity is the Perimeter. For agents, that means a unique identity, a named human owner, scoped access, and a log tied to that identity.
Every agent. Accountable.
If you run AI in a regulated firm: could you shut off one specific agent today without breaking the others?