Okta and the Urgency of an "Off Switch" for AI Agents

Todd McKinnon, CEO of Okta, recently emphasized the need for companies to integrate a "kill switch" into their artificial intelligence (AI) agents. This measure would allow for the instant revocation of these agents' access in case of issues, thereby ensuring enhanced security. McKinnon explained that companies want to be able to react quickly if something goes wrong.

Okta, which positions itself as a security layer for these agents, is already used by millions of workers to access various applications and services. With the rise of AI agents, which may soon require similar access, McKinnon explained that his company is focused on securing these interactions. He advocated for a kill switch to control AI agents when things go awry.

In an interview on The Verge's "Decoder" podcast, McKinnon described these agents as a new category of digital workers. Capable of accessing systems, manipulating data, and performing actions within a company's software, these agents require strict controls. McKinnon stressed the need to monitor these agents, define their roles and permissions, and have the ability to "unplug" them if they become uncontrollable.

Companies are increasingly experimenting with AI agents to automate workflows, develop software, and even manage certain physical tasks. This type of power requires strict parameters, he stated. The benefits of AI agents include productivity, but the risks are equally significant and require thoughtful limits.

Okta's proposal is not aimed at eliminating AI agents but at limiting their access to sensitive data. Harish Pari, senior vice president of AI security at Okta, highlighted that the risk is already present, as AI agents need to access sensitive systems and data, thus creating new attack vectors.

Okta issued a press release on March 15 titled "The Plan for Secure Agentic Enterprise," stating that companies must be able to instantly revoke agents' access to all systems to contain risks. This framework also includes real-time enforcement of data-sharing permissions, human approval for risky actions, and detailed audit logs that track every decision and access attempt by the agents.

Meanwhile, a proposed California bill by Senator Scott Wiener, which required companies to integrate a security system for AI, was supported by Elon Musk but vetoed by Governor Gavin Newsom. Despite this, McKinnon insisted on the importance for private companies to develop their own security systems, emphasizing that things will go wrong, and there will be problems, threats, and prompt injections. It is crucial to prevent agents from accessing data in a crisis, almost as if you were unplugging a machine from the network.