AI Coding Weekly

Ubl says Cursor leak proves static credentials must die

The secrets did not leak from a vault, they leaked from saved model conversations.

Cursor published an update on its recent security incident, promising a detailed overview of what happened along with its remediation steps.

Malte Ubl read it and reduced it to one line. His summary is that static credentials leaked through stolen LLM transcripts that contained the credentials. That is the whole shape of the incident as he describes it. The secrets were not exfiltrated from a secrets manager or a repo, they were sitting inside saved conversations with a model, and whoever got the transcripts got the keys.

Malte Ubl
@cramforce
X
The only real fix is to kill long-lived static credentials.
Nov 14, 2026 · View on X
Malte Ubl
@cramforce
X
static credentials leaked through stolen LLM transcripts that contained the credentials
Nov 14, 2026 · View on X

Ubl's conclusion is not a Cursor-specific one. He says the only real fix is to kill long-lived static credentials, and that this is why he published a post arguing for it, adding that it is time for the industry to come together around this.

Why the argument lands harder now

The case against long-lived static credentials is old and uncontroversial in principle. Short-lived, automatically rotated credentials limit the blast radius of any leak, because a stolen token that expires in minutes is worth far less than one that works for a year.

What the incident adds, on Ubl's reading, is a new place for secrets to accumulate. A transcript is a durable artifact of an agent session, and it is held by whoever runs the agent rather than by you. Anything a developer pastes into a chat to unblock an agent, a connection string, an API key, a token copied out of a dashboard, ends up stored somewhere outside the systems built to protect secrets. Every credential that outlives the session it was pasted into is a liability that sits there until somebody takes the logs.

One caveat on attribution. Ubl is offering a summary and a position, not an incident report. The specifics of what happened and what Cursor did about it are in Cursor's own post, and his tweet does not restate them. If you run agents against anything with production access, that post is the one to read, and his argument is the one to have with your platform team afterwards.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.