Identity management for agentic AI
A version of this post appears on the Web Frolic blog.
It's possible I'm missing something in the big picture, but I strongly feel there's a lot of reinventing words to make this look like a new problem.
The concepts already exist
- An agent acting directly: we already have service accounts.
- An agent acting on behalf of a user: we already have access delegation.
Where everyone falls short
There are three core areas where most of us have fallen short and managed to get by until now.
1. Auditing. Who did what? For most applications the user acted directly, so delegated access wasn't a concern.
2. Authorization controls. In most applications, client-side UI controls hid many of the actions a user could technically have executed. Agents explore and behave dynamically, so those controls don't hold anymore. A lot of this is now being labeled as agents "hacking."
3. Granular access controls. Many applications that supported delegated access did it the easiest way possible: a user API key or token with the same permissions as the user.
Giving an agent full access and crossing your fingers that it won't do anything else isn't a security control. It's a wish.
Missing tooling
Another big challenge is that many applications have no tooling for this at all on the service side. So the agentic ecosystem works with what exists, using the user's regular credentials to act on the user's behalf. To me that's an own goal for both the agent and the service provider, and we shouldn't encourage it.
How ready is your identity provider?
Is it giving you the tools to prepare for agents and welcome them?
If you're looking for better options, reach out for an early demo 😀 My email and LinkedIn are on the About page.