← Back to Feed

Post from 8 October 2026

An abandoned AI project can leave behind something you probably don’t want: an identity that still has access to your production systems.

Imagine a team experimenting with an internal AI agent that helps support engineers investigate customer issues. They give it a service identity with read access to production customer data.

The prototype works, but priorities change. The project gets abandoned, and everyone moves on.

Six months later, the service identity still exists, with the same production permissions. If its credentials leaked through an old CI/CD pipeline or a compromised development environment, an attacker could use them to read customer data. From the system’s perspective, those requests come from a legitimate identity with valid permissions. And because nobody is actively maintaining that project anymore, who is monitoring that identity’s activity? Who is reviewing whether it still needs access?

Now multiply that across dozens of teams experimenting with internal AI agents, each creating identities and granting permissions to different systems. You could end up with hundreds of identities that nobody uses or reviews, all still holding valid permissions.

This made me realize there’s a hidden assumption in how we manage identity lifecycles: something will tell us when an identity is no longer needed.

For employees, we have HR triggers. Someone leaves the company, and an offboarding process revokes their access. An agent doesn’t resign. Its project might simply get abandoned. So what triggers that process for an agent?

This isn’t a new security problem. We’ve dealt with forgotten service accounts for years. But as organizations experiment with more internal AI agents, I think we’re going to see more of it unless we design for it.

A few things I’d want in place:

  • Every agent identity has a designated owner accountable for its access.
  • Temporary identities expire unless someone explicitly renews their authorization.
  • Retiring a project automatically triggers revocation of its associated identities and permissions.

Short-lived access tokens help, but they’re not sufficient if the underlying identity can keep obtaining new tokens indefinitely.

So when reviewing the architecture of an internal AI agent, I’d add one more questions:

“What will cause this agent’s access to be revoked when nobody needs it anymore?”

Also on Bluesky · Mastodon

Loading comments…