Frontier models are rentals, so vault the keys that rent them
Nov 16, 2026
Frontier models are rented, and an API key controls the account. Those keys pile up in repos, chat threads, and spreadsheets, and AI agents often run with more access than the person who launched them, or less than the task needs. An encrypted vault, separate records for human and agent credentials, approval before release on production keys, and a log of every access give security teams control over who uses each key.
Nobody owns a frontier model. You rent it with an API key and pay by the token, and the key goes wherever the last person who needed it pasted it: a .env file, a Slack thread, a shared spreadsheet, a teammate's browser.
Whoever holds the key is the account
They can spend your budget, send your prompts and data to the provider, and act as your application. Provider usage reports track the key, not the person who copied it out of a shared note.
Then the pilots multiply. Each one gets its own key, each team tries a different model, and contractors need temporary access. When an engineer leaves, nobody can say which keys they touched.
Agents shouldn't borrow a person's login
Run an agent on a developer's personal login and it inherits everything that developer can reach, production systems included, whether the task needs them or not. Hand it a shared admin key and it can reach more than the developer who launched it. In both cases the access has drifted away from the job.
Store human credentials and agent credentials as separate records, with separate roles and a named owner on each. Agent and model keys get their own permissions and their own audit trail, so you can answer "which systems can this agent reach" from the record.
- One key per agent or application. No sharing across projects.
- Scope each key at the provider to the least access the task needs.
- Give each agent its own key, stored in its own record, so it never runs on a person's login.
- Name an owner for every key.
Put the keys in a vault with a log
Keep model keys in a centralized, encrypted vault. Use role-based access controls to decide who sees or uses each record, so a developer on one project doesn't see every key in the company. For production keys, require a set number of approvals before the password is released. Enable MFA for user logins.
Log every access: who opened the key, when, and what they did. If a key leaks or someone leaves, you start from that record.
Host the vault somewhere your team controls.
Start with an inventory
Search your repos, wikis, and chat history for model provider keys. Move what you find into the vault, split human and agent records, and assign owners.
Meet Password Secure
Once the keys are inventoried, they need a home with controls around it. Netwrix Password Secure is a self-hosted password management solution that stores them in an encrypted vault, limits each record to the roles that need it, adds an approval requirement for production keys through seals, and logs every access. You run it on premises or in your own cloud infrastructure, so the keys stay in an environment your team controls.
Request a demo to see how Password Secure handles the credentials behind your AI tools.
See it in actionShare on
Learn More
About the author
Sascha Martens
Chief Technology Officer
Insights from a security professional dedicated to breaking down today’s challenges and guiding teams to protect identities and data.