LAS VEGAS — "No one company can secure AI agents alone," said Okta CEO and co-founder Todd McKinnon during his keynote address at the Oktane conference here Wednesday (Sept. 23).
That was a refreshing statement to hear from the head of a major company in the cybersecurity industry. Dozens of other vendors claim that they, and they alone, possess the magic bullet that will solve every client's worries about
AI governance and security.
"Our biggest competitor is confusion," McKinnon said, because customers don't know where to turn to control the AI agents they're rapidly deploying. "Everyone says they've got the solution for managing AI."
Collaboration vs. competition
Okta admits that it doesn't have all the answers. The company does offer some pretty robust solutions for controlling AI-agent authentication, authorization and governance, but as Okta President and COO Eric Kelleher said in a Q&A session after McKinnon's keynote, "It's going to take collaboration at every part of the stack."
Click here for more coverage from this year's Oktane conference.Identity providers like Okta can handle the identity part of the technology stack, Kelleher added. Network-security providers will have to come up with solutions for the networking aspects of AI security, endpoint-security providers the endpoint parts, and so on.
To that end, Okta this week announced the
Blueprint Alliance, a consortium of companies from different parts of the technology industry pledging to work together to devise common, open standards for AI security.
The founding members of the Blueprint Alliance include CrowdStrike, Google Cloud, Amazon Web Services, Salesforce, Wiz, ServiceNow and Proofpoint.
None of the frontier AI developers like OpenAI or Anthropic have signed on, but Kelleher hinted that at least one may join soon. (Anthropic's very visible presence at Oktane offered a clue.)
McKinnon said that the Blueprint Alliance seeks to devise open standards that will answer four basic questions: Where are my agents? What can they do? What are they doing? And how do I respond?
"The technology around AI is too complex," said McKinnon. "We need standardization on a massive scale."
Setting standards
Okta has had experience in forming open standards, as last year it launched what it calls
Cross App Access, an extension to the OAuth standard that places AI agent interactions with MCP servers, applications, and other resources under the control of identity-management systems.
Cross App Access is being slowly adopted across the industry, but in his keynote, McKinnon called on one of the creators of an even more widely used AI open standard — Anthropic's David Soria Parra, who worked on putting together the now-ubiquitous
Model Context Protocol.
"Why was it important that MCP was developed as an open standard?" McKinnon asked Soria Parra.
"We want to build it together as an industry," Soria Parra replied.
Okta is extending Cross App Access to the modern single-sign-on protocol that Okta helped develop for human users years ago, enabling AI agents to sign on once and log into many services as part of a service called Agent SSO.
McKinnon and Mallory Sword Glenn, Director of Okta for AI Agents Product Marketing, ran through an on-stage demonstration in which Sword Glenn manually granted a Claude Code instance access to her Atlassian, GitHub and Slack accounts, logging into each one by one, so that the agent could generate a report and send it to McKinnon.
"You can see that this is very tedious and very insecure," she said.
But after activating Agent SSO, Sword Glenn logged into Claude Code again and the connections to her accounts were already made as part of the SSO process.
"For me as an end user, this is great," she said. "For my IT and security teams, it's phenomenal," because the agent logins are monitored and controlled.
"So why hasn't anyone done this before?" McKinnon asked the crowd. "Because it requires standardization, and that's a lot of work. We've been working on it for four years."
A gateway and a kill switch
Only by implementing such common standards, he said, can companies move forward with rapid AI innovation confident that they're not risking too much security at the same time. The choice between innovation and security is often misrepresented, McKinnon said.
"Either lock it all down and don't connect to everything, or let it connect to everything," he said. "Not a great choice."
But there's a simple solution, he added: "Visibility and control. That's how we break this deadlock."
That's why Okta is adding more features to its
Okta for AI Agents product, which was itself introduced at last year's Oktane and became generally available this past April.
One is Shadow AI Agent Discovery for Endpoints, which can detect "shadow AI" agents running on employee endpoints, and which can then be brought under the control of Okta through Okta for AI Agents. It will be available this quarter.
Another is support for third-party identity providers. In other words, you don't need to use Okta as your primary identity-management solution in order to use Okta for AI Agents.
"We want to support organizations that use other IdPs for humans," McKinnon explained. "The agents can now authorize with your existing identity provider."
There's also a Configuration Designer that maps out agent-to-resource connections so that they can be managed and audited, available soon.
Finally, the Agent Gateway, a previously announced MCP server built into Okta for AI Agents that governs agent connections to resources, should be generally available soon. And by the end of 2026, the Agent Gateway will get a kill switch that can immediately revoke all tokens and cut off resource access for an agent that's gone rogue.
"Our vision is to free anyone to safely use AI technology," said McKinnon. "We have to bring visibility and control to this industry right now."
