Agentic AI security pushes vendors toward shared safeguards

Okta is making a simple case: agentic AI will not be safe if every vendor builds its own little fortress. As autonomous systems start moving across products and platforms, the company says security has to be coordinated, not siloed. That message came through at Oktane, where Charlotte Wylie, Okta’s senior vice president and deputy chief security officer, argued that trust in agents will depend on shared safeguards that work across the stack.

The company’s answer is the Blueprint Alliance, which includes founding members Amazon Web Services, CrowdStrike, Salesforce and ServiceNow. The group is trying to make security controls and risk signals interoperable across agentic systems. Wylie’s point was blunt: vendors can’t solve this one alone, because customers are moving too fast and the technology is already spreading too widely for isolated controls to keep up.

The architecture she described starts with basic discipline. Each agent needs an identity. Its permissions need to be limited. Its behavior needs to be visible. Put those together, and an organization can spot when an agent drifts away from what it was supposed to do. That is the whole bet here: not just trusting an agent, but constraining it enough to catch it when it goes rogue.

Wylie said products like Okta for AI Agents can help govern permissions and monitor what those agents are actually doing. The goal is to stop a bad action before it spreads, not after the fact. And she also pushed the industry to explain this in plain language, outside the usual security bubble, so the people using these systems at work and at home understand how they’re being handled.

This is not a flashy AI story. It is a boring one, which is usually a good sign in security. Shared rules, shared signals, fewer solo heroics. That’s probably the only adult way to handle software that can act on its own and wander around other people’s systems like it owns the place.