Skip to main content

MFA without a Human - AI Security Essentials briefing #3

  • Nov 9, 2026 10:17 PM UTC
  • 0 replies
  • 9 views

Forum|alt.badge.img+3

Read the quick briefing from our third live virtual AI series session here, or watch the recording of the full conversation below (at the bottom of the write up). We are continuing to deliver this series, so sign up for upcoming sessions here.

 

Briefing recap

The problem 

Machine identities already outnumber people 82 to 1 in the typical enterprise, and only 22% of teams treat AI agents as independent identities. Token theft is replacing password theft because tokens leave fewer traces. Every MFA factor assumes a person supplies it, and an agent can't: no phone, no biometric, no one to approve a push. What's left is a static secret, and whoever holds it is the machine. 

 

Why the gap is wider than it looks 

NIST's own guidelines separate verifying an identity from authenticating with a credential. For a person, someone checks a government ID first. For a machine, that step usually never happens: a service account gets created and nothing verifies what will run under it. It's the same gap that let one stolen signing key forge tokens for roughly 60,000 emails in 2023. The deterrents that keep people honest don't apply to an agent, so trust has to come from the credential itself. 

 

What this means for Delinea customers

Three things worth addressing this week: 

  • Find every agent running on a borrowed credential. A developer's token or session reused by an agent is the most common failure pattern today, and it fails every identity check at once. 
  • Move federation-capable workloads off static secrets. A workload that already has an identity with its provider can sign in with a short-lived, scoped token instead of a static one. Nothing left to vault or rotate, and the exposure window shrinks to that token's lifetime. 
  • Name a human sponsor for every agent. The standards don't require this yet, so it has to be policy. An unsponsored agent is the machine equivalent of an orphaned service account. 

 

Where to start 

We recommend starting in this order: 

  1. Find every agent running on a borrowed human credential, since that's the pattern that fails every identity check at once.
  2. Move federation-capable workloads off static secrets and retire their client secrets.
  3. Turn on attribution for agents on endpoints and write one policy: no unverified agent in production, with a named owner for every agent you run. 

Check out the recording of the full conversation here.

Sources included: Microsoft State of Multicloud Security Risk Report (via Delinea blog); CSA research note on NIST IR 8587 (Sept 2026); CISA CSRB review of the 2023 cloud email intrusion; Delinea Labs, Threat Outlook, Dec 2025; NIST SP 800-63-4 (July 2025).