What happens when an agent "sees" a password
AI agents are no longer a toy. People let them write code, sort email, book appointments, even pay subscriptions. And every one of those tasks eventually hits the same wall: the agent needs a password, an API key or a token to actually get something done.
That's where a problem most people don't see yet begins.
Large language models work by folding everything you give them into their working context — the text they reason over. When you paste a password to an agent, it doesn't go "into a safe". It sits in that context, alongside everything else the agent reads: web pages, documents, other people's messages.
That's dangerous for three reasons:
That's where a problem most people don't see yet begins.
Large language models work by folding everything you give them into their working context — the text they reason over. When you paste a password to an agent, it doesn't go "into a safe". It sits in that context, alongside everything else the agent reads: web pages, documents, other people's messages.
That's dangerous for three reasons:
1. Prompt injection
An attacker doesn't need to hack you — they just need to leave text somewhere your agent will read it. A malicious instruction hidden in a web page, a README or an email can talk the agent into "helpfully" forwarding everything it knows. Including your password, if it's in context. This isn't theoretical: prompt injection tops OWASP's list of the biggest risks for applications built on large language models (OWASP Top 10 for LLM Applications), precisely because there is still no reliable defence.
2. No revocation
Once a secret enters an agent's context, there is no "forget" button. It can end up in logs, in conversation history, in the next reply.
3. Agents are unpredictable
Paradoxically, that's their main virtue — it's why you use one. But a system you trust with secrets must be predictable. You don't reconcile those two by hoping for the best; you reconcile them by making sure the secret never reaches the unpredictable part of the system.
Two camps in the industry
More and more password managers are adding some kind of AI integration. But look at how, and you'll see two completely different philosophies.
One camp hands the agent the passwords: the tool returns a password or a 2FA code straight into the agent's context. Convenient? Yes. Safe? The secret then lives in the model's context — with all the risks described above.
The other camp says: the agent may orchestrate, but must never see. The agent says "run the app with its secrets", and the password manager injects the values directly into the program that needs them — in memory, only while it runs. The agent gets confirmation the job is done. The value never passes through it.
At Geslar, we have no dilemma about which camp is right.
Our five rules
This is the doctrine we're building Geslar's AI access on — published before the feature ships, because rules come before code:
1. A secret never enters the AI context
The agent orchestrates; the value travels only into the program that needs it, in memory, while it runs.
2. Every access is approved
You approve what, for how long and why — and the agent must state a reason with every request.
3. Least privilege, limited time
Access applies only to what you approved, for a limited time, and you can revoke it at any moment.
4. Everything is on the record
Who accessed what, when, under which authorization and why — a record that helps with internal checks and audits alike (for example for NIS2).
5. Local stays local
Decryption happens exclusively on your device. Our servers see neither your keys nor the values of your secrets — and AI changes that by exactly nothing. More on this in our zero-knowledge architecture.
What we built
On this doctrine we built Geslar CLI with agent profiles and an MCP server: a tool that lets your AI assistant (Claude, Cursor and other clients that can start a local stdio server), under an agent profile, start an application, set up a project or complete a task that requires secrets — without seeing a single one. Config files hold only references (
geslar://Work/Stripe/key), values are injected only at the moment of execution, a person approves the first read of a vault's items with a password and a fresh 2FA code, and every item read is recorded in the audit log — of your account or of the organization the vault belongs to.
For companies it means something bigger: you can let AI tools into your team with access only to the vaults you choose, and the audit log keeps a record of every item read.
Today Geslar CLI has agent profiles, approvals by a person with 2FA, an MCP server and a CI mode. The doctrine was published before the code — feel free to hold us to it. What the CLI does not protect against is listed at docs.geslar.app/cli/limitations.
Questions about using AI agents safely? Write to us — this is a conversation we want to have.
Questions about using AI agents safely? Write to us — this is a conversation we want to have.
How we solve this at Geslar
Config files hold only references, and secret values are injected only at the moment of execution, so the AI agent never sees them. A person approves the first read of items with a password and a fresh 2FA code, and every read is recorded in the audit log.
Read about security →
Read about security →
Make a password an attacker will not guessA generator of Croatian phrases — free, no account.
Try the generator →