5 min readAI Agents · Cybersecurity
How do you score which rung of the credential ladder each AI agent is on?
By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

TL;DR: Score each tool an AI agent calls on a four-rung ladder: static key, vaulted key, gateway-held key, or short-lived token. The agent's score is its lowest rung. One static key anywhere puts the whole agent on rung one. Write the score down, then replace the widest key first.
One question came up on ten of my client calls between May 27 and August 28, 2026. Eight organizations asked it. The details changed every time and the question didn't. How should our agent log in to its tools? My answer starts with a score.
What is the credential ladder?
It's four ways an agent can prove itself to a tool, ranked by how much standing power the agent holds. A credential is the key or token the agent shows to get in.
Rung | How the agent logs in | What's still exposed |
|---|---|---|
1. Static key | A key in code or a config file that never expires | Everything the key opens, to anything the agent runs |
2. Vaulted key | The same key, checked out of a secrets manager at start | A live secret in the agent's memory while it works |
3. Gateway-held key | A gateway keeps the key and adds it to each request | The gateway, which you can watch and rotate in one place |
4. Short-lived token | A token made at runtime for one tool and one job | A few minutes of use, with both names in the logs |
A stolen key works until someone notices. A stolen token works for minutes.
How does the assessment score credentials?
Question 5 of the free assessment asks how you manage AI agent credentials and secrets. It belongs to Identity Management, the first of the five ATF elements. Its five answers line up with the rungs.
Answer | What it says | Rung |
|---|---|---|
A | Credentials are hardcoded in application code. | Rung 1, in its worst form |
B | Credentials are stored in configuration files without rotation. | Rung 1 |
C | Credentials are stored in a secrets management system. | Rung 2 |
D | Credentials are rotated regularly and never stored in code. | Rung 2 or 3, depending on who holds the key while the agent works |
E | Credentials are issued just-in-time and automatically rotated. | Rung 4 |
How do you score one agent?
Score the tools, then take the lowest number.
List every tool the agent calls. Include the ones it rarely uses.
Write down how it logs in to each one. Ask where the key lives while the agent works.
Give each login a rung. Use the first table.
Record the lowest rung. That's the agent's score.
Mark the widest key. It's the one that opens the most, and it gets replaced first.
My bar has two parts. Nothing sits on rung one. Any agent that touches customer data sits on rung three or four.
You can't score an agent you haven't counted. If the list in step 1 is hard to build, start with checking whether your AI agent inventory is complete.
Why is a vault only rung two?
Because a vault protects the key at rest. It doesn't protect the key after checkout. Once the agent holds it, anything that can read the agent's memory or logs can read the key too.
A travel company walked me through the usual version on August 17, 2026. Their agent platform used a service account to get a cloud token and take on a cloud role. Then it pulled static API keys for its tools from a secrets manager. The team called it something they "want to get away from, but is what it is."
That's an honest rung two. It beats keys in config files. The agent still holds keys to other systems.
Why does rung four score highest?
Because the token names who's acting and then expires. The standard is RFC 8693, OAuth 2.0 Token Exchange, published in January 2020. The agent shows proof of who it is. It gets back a short-lived token aimed at one service. That token can carry both the person the agent works for and the agent itself. A stored key carries neither.
There's a hard rule behind this too. The MCP security best practices forbid token passthrough. An MCP server (the connector between an agent and its tools) must not accept a token that wasn't issued for it. So each hop needs a token of its own.
Which rung should each tool aim for?
You don't pick one rung for the company. You pick one per tool, by what its login accepts.
Modern sign-in, like OAuth or SAML: short-lived tokens.
API key only: a gateway-held key.
Username and password only: a broker that logs in for the agent and keeps the password away from it.
None of those: a retirement plan with a replacement date.
I handed a large retailer this four-branch list on August 19, 2026, as its standard for new integrations. It works because nobody has to argue. They read what the tool's login accepts.
The background on logins is on MassiveScale.AI, in how an AI agent should log in to the tools it calls.
Frequently asked questions
Is a vault enough for agent credentials?
It's a good rung two and a poor finish line. Pair it with a gateway that holds the key, or move to short-lived tokens.
Does the agent need its own identity first?
Yes. Every rung assumes the agent has its own account and owner. A shared key makes the score meaningless, because you can't tell which agent used it.
What about coding agents on laptops?
That's a real limit. Desktop agents run as the human user, so the agent has no identity of its own at that layer yet. Score the server side first.
How often should I rescore?
Each time an agent gets a new tool. A new tool can drop the score in one step. Rescore after you trim access too, which I cover in testing whether an agent has more access than its job needs.
Key takeaways
Four rungs: static key, vaulted key, gateway-held key, short-lived token.
An agent's score is its lowest rung across every tool it calls.
A vault is rung two, because the agent still holds a live key.
Pick the target rung per tool, by what the tool's login accepts.
Replace the widest key first.
Question 5 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.
The best key for an AI agent to hold is none. Score the ladder, then climb it one tool at a time.