San Francisco — The problem with giving an AI agent a key to the company network has always been simple: that key never changes. Traditional API keys sit in code, in configuration files, often for years. An agent that steals one can act as that agent forever.
Okta’s Cross App Access framework, announced last June, was built to kill that static-key model. This week, the company disclosed that more than 25 software vendors have now signed on to use it.
The early adopters list reads like a who’s who of enterprise software. Asana, Atlassian, Cloudflare, Datadog, Slack, Zoom. Anthropic’s Claude is in there too — one AI company letting its agents talk to other apps through Okta’s controls rather than through hardcoded tokens.
The framework sits as an extension of OAuth, the standard that already governs how users grant apps limited access to their data. Okta’s version, Cross App Access or XAA, applies that same logic to machine-to-machine connections.
What changes for the companies that adopt it? An agent that wants to pull data from Salesforce or post a message in Slack no longer carries a permanent credential. It has to ask for permission each time, or at least within a window set by the company’s identity policies.
If an agent is compromised, the damage stops at the next expiration. That is the theory, anyway. The practical consequence is that security teams get visibility they did not have.
Most organizations today have little idea how many agent-to-app connections exist inside their networks. The connections are invisible, buried in CI/CD pipelines or background scripts.
Okta’s framework routes all that traffic through a central identity layer. If an agent connects to a database, the security team sees it. If an agent that should only read customer data suddenly tries to write, the system can block it.
This matters because agent adoption is accelerating faster than security controls can keep up. The report notes that most agent connections still depend on static API keys.
That is a mismatch. Companies are pushing agents into production without the guardrails that have existed for human users for years. Human employees get single sign-on, multi-factor authentication, session timeouts.
Agents get a key that never expires. The framework now serves as an official authorization extension for the Model Context Protocol, the standard that connects AI models to outside data and tools. That is a structural move.
It means any company building on MCP — and many are — can plug into XAA without building custom identity logic. The integration is baked into the protocol itself.
The partner list includes Canva, Docker, Figma, Linear Orbit, Supabase, and developer tools Cursor and Microsoft’s Visual Studio Code. Those are not just the usual enterprise suspects. They include the tools where developers actually build agents.
That is where the credential problem starts. A developer writes an agent in VS Code, tests it with a hardcoded key, then pushes it to production.
The key stays hardcoded. XAA aims to interrupt that pipeline by giving developers a way to issue temporary, scoped credentials instead. What happens next depends on adoption.
Twenty-five vendors is a start, but enterprise security is only as strong as the weakest connection in the chain. If one major application refuses to support the framework, agents will still use static keys to reach it. The framework is open and built on existing standards, which lowers the barrier for others to join.
Whether they do is the open question. The problem Okta is solving is real.
The question is whether the industry moves fast enough to close the gap before a major agent breach makes the point for them.
Sources
- Okta advances the industry standard for secure AI agent connections…
- Okta expands Cross App Access to secure AI agent enterprise…
