All notes

Agents run on your keys. Attackers just proved it

Two things happened in the last two days that belong in the same sentence. On October 21, Railway posted its security disclosure: an unauthorized third party may have reached the systems that store customer environment variables and, for some of them, obtained the decrypted contents. Environment variables are where deployed apps keep their API keys. Railway believes the attack was part of a series targeting platform companies with the aim of compromising AI inference providers, and it told every customer to rotate everything.

One day later, 1Password shipped 1Password for Agents: the agent proves who it is, you approve the sign-in on your device, and 1Password fills the password without the agent ever seeing it.

Put those next to each other and you can see where the industry is being pushed. I made the case a while back that agentic AI needs a permission interface, not a better prompt. This was the week the argument stopped being theoretical.

On this page

Why are attackers stealing AI API keys?

Because a stolen AI key is preloaded with spend. It is a credit card somebody else already paid to open, and the bill arrives in your name. Railway's disclosure says the goal was compromising AI inference providers — the platforms running the models your app calls. Unlike a database dump, an AI key converts to money automatically, every request, until someone notices.

Why is a plaintext key in agent context the wrong model?

Most of us hand agents credentials by pasting a key into a prompt, an env file, or a config the agent reads. The moment you do, that key is visible to the model, to its session memory, and to any log that ever recorded the exchange. If the agent is compromised, the credentials go with it. This is the pattern 1Password's own developer documentation is built to kill: secrets should be resolved at the moment of use, not carried inside the thing you cannot fully audit.

The reframe I keep coming back to: a credential is not something the agent should hold. It is an authorization event. The agent should do what a real person does — prove identity, request access, get stopped when it should not proceed. Holding a standing copy of the password is what makes an agent dangerous, and standing permission is exactly what we should be signing away.

What does 1Password for Agents actually change?

The flow is three-party. The agent authenticates to 1Password with a digital key. You get a notification naming the account and the reason for the request. If you approve, 1Password fills the sign-in. The agent finishes the task with proof it was you, and never holds the secret it needed.

The per-request approval matters more than the autofill. Approval is scoped, visible, and revocable — the shape a permission interface should have. It shipped with support for agents including Claude Duet, ChatGPT Work, and Grok Bot. Is it the final design? Probably not. But it is the first mainstream product where giving an agent access and giving it a password are different things, and that is the distinction the rest of the stack has to catch up to.

What should you fix today?

Railway's advice applies regardless: roll every secret stored on the platform, registry credentials included, and check the logs for unusual activity. Do it even if you read about the breach a week from now — the cost of a rotation you did not need is minutes, and the cost of the one you skipped is your bill and your data.

Beyond the immediate cleanup, three habits hold up. One key per app, so a compromise stays contained to one product. A spend cap on every agent's provider key, so the damage is bounded even when rotation is slow. And a standing refusal to hand anything long-lived to an agent directly. If you cannot revoke your agent's access from your phone in under a minute, you gave it too much.

The uncomfortable takeaway from this week: the interesting attack surface stopped being your code and moved to your credentials. Agents are where your product does its work, so they are where your risk lives. Design accordingly.

Frequently asked questions

  • On October 21, 2026, Railway detected anomalous activity in the systems that store customer environment variables. Its disclosure says an unauthorized third party may have accessed encrypted variables and obtained the decrypted contents of some of them, which can include API keys and database credentials. Railway paused deployments and told customers to rotate every secret stored on the platform.

About the author

mosh

mosh is a product designer for growth, working with design thinking and ever-improving design systems. What matters: fixing conversion, whether in B2B dashboards or direct-consumer apps.

Keep reading