AI Agent Identity, Limits and Kill Switches: Walmart to OpenAI

On October 21, Walmart described a problem almost every company has: a customer's AI agent signs in as the customer, and nobody can tell the two apart. Walmart's fix has three parts. Each agent gets its own identity, its own limits, and a kill switch the customer controls. Most companies cannot yet say which agents are running in their own systems this morning. This video covers the three ways agents get in; what shared identities and standing trust have cost this year, from Zenity's Bedrock AgentCore research to Microsoft's report on Storm-0940; one agent, one identity, one owner; the six limits every agent needs; why OpenAI's alarm worked and its switch did not, and what a working kill switch looks like; Walmart's, Amazon's and American Express's three answers to outside agents; what the Fed, the OCC and the FDIC are asking banks; and a ninety-day plan. 0:00 The agents you cannot see 1:49 Three ways in 3:02 Shared identities, standing trust 5:00 One agent, one identity, one owner 7:09 Limits 9:11 The kill switch 11:19 Outside agents at the front door 13:16 The regulators are asking 14:43 What it's worth, and what to do The full article, with every source linked: https://roai.us/insights/which-ai-age... The research paper, Who Let That Agent In?, with a register template for every agent you run, the six limits to set on each, a kill switch drill you can run in an afternoon, a policy for outside agents and a page of answers for your examiner: https://roai.us/research/who-let-that... Named in this video: Walmart, G2, the Cloud Security Alliance, Palo Alto Networks, Microsoft, Zenity, Amazon, ConnectWise, OpenCode, OpenAI, Anthropic, Hugging Face, Salesloft, Meta, American Express, Perplexity, Reuters, the Federal Reserve, the OCC, the FDIC, the Five Eyes cyber agencies, NIST, the EU and Singapore's central bank.