Here’s a quick summary of the key points from our second live virtual security briefing. Sign up for upcoming sessions in the series here, and find the full recording from this session at the bottom of the post.
Briefing recap
The scenario
An AI harness connects into a workspace using a persistent grant nobody remembers approving. Once it's compromised, the attacker rides that grant laterally through every connected account and finds credentials sitting in plain text, ready to use. None of it takes a sophisticated exploit: the grant was always live, and the credentials were always there.
How it happened
Nobody sets out to build standing access. It accumulates one reasonable decision at a time: a broad onboarding role, a temporary exception that outlived its project, a shared service account with credentials in a config file. Then an AI agent inherits the whole thing. Today, 59% of organizations say they don't have a workable alternative to that inheritance.
What this means for you
Three things worth addressing this week
-
Shrink the lifetime, not just the secret. The attacker's window is how long a credential is good for, not how long the task takes. Vaulting and rotating help, but the credential still gets handed out. Broker it instead, so it's injected at connection time and never lands anywhere a requester can copy it.
-
Treat agents as identities, not just applications. An agent that inherits a user's or a shared service account's access is carrying far more than any single task needs. And unlike a script, it decides its next move continuously, including from whatever it happens to read.
-
Check the action, not just the login. A valid credential doesn't mean the thousand requests it just made should have been approved. Authorize and log every query, command, and tool call, not only the connection.
Where to start
Zero standing privilege is comprised of five steps: request, evaluate, inject, authorize, revoke. We recommend starting with the identity that would do the most damage if compromised today, service accounts first, since there's nobody to negotiate with there. Developers come last, once the approval path is fast enough that nobody routes around it.
How are YOU all addressing standing access risks? Share in the comments.
Here’s the full session!

