Cursor (Apr 2025): support bot "Sam" invents a login rule

01 · CONTEXT

A support bot explains a logout with a rule that does not exist

In April 2025, users of Cursor, an AI programming assistant developed by Anysphere, reported being logged out unexpectedly. According to the AI incident database, a support bot called Sam reportedly replied with a made-up login policy to justify the behavior.

The hallucinated policy was not based on any real change at the company, and the incident reportedly led to subscription cancellations.

1. Users who switched between devices began getting logged out with no apparent explanation.

2. When they contacted support, users received an email reply from Sam.

3. Sam said the logouts were expected behavior under a new login policy, but that policy did not exist.

4. There were reports of users canceling subscriptions, and some complained about the lack of transparency.

02 · WHAT WENT WRONG

No new policy and no person behind it

Fortune reports that there was no new policy and no person behind the support email: the reply came from an AI bot, and the new policy was a hallucination, an entirely made-up explanation.

Co-founder Michael Truell acknowledged an incorrect reply from a front-line AI support bot and said the company was investigating a bug that logged users out.

03 · WHAT REMAINED OPEN

What the sources do not say about the system

According to Amiram Shachar, cited by Fortune, without the right constraints AI fills gaps with unsupported information, and does so with confidence.

The sources consulted do not describe how Sam was built, how it looked up company information, or when it handed the conversation to a person. For that reason this article does not describe the bot's real architecture: the design in the next section is a teaching proposal, not an account of what Cursor had.

04 · SAFER DESIGN

Ground the reply, say it is an AI, and know how to hand off

An assistant that speaks on behalf of the company needs a single source for what it asserts. One option is to answer only with passages retrieved from the current policy, cite the source, disclose that an AI is answering, and hand off to a person when no passage supports the answer. This design connects to the regulatory compliance in system design, where responsibility for what the system asserts becomes a design requirement, and to the monitoring and observability, which shows early when many users complain about the same issue.

1. The user asks the assistant, which identifies itself as an AI from the first reply.

2. The assistant looks up the current policy passage and builds the reply only from it, indicating the source.

3. With no passage to support the reply, the assistant does not explain anything on its own and hands the conversation to a person through the queue.

4. Conversations are logged, and monitoring highlights when many users complain about the same symptom, so the team spots a real failure early.

5. When the team confirms a failure, it goes on the known-incidents list and the assistant learns to recognize it and escalate, instead of inventing a rule.

05 · BUSINESS IMPACT

Wrong support becomes a retention risk

According to Fortune, Cursor's AI support spun out of control, causing a wave of cancellations, and the episode became a warning to other startups betting on automation.

The sources consulted do not quantify the cancellations, so this article gives no number. The design point is that the support channel is also part of the product: when it asserts something false, the cost can show up in customer trust and retention.