Cursor's AI support agent invented a login policy, and users said they cancelled

April 2025 — A coding-tool startup let an AI model answer its support email, and the model made up a rule. Here's what launched, what it cost, and what went wrong.

Cost and outcome (as of April 19, 2025)

What the bot said: "Cursor is designed to work with one device per subscription as a core security feature." (Ars Technica)
Customer impact: public cancellation posts on Reddit; Fortune reported "a wave of cancellations." No count was published. (Fortune)
Remedy: the user was refunded, and AI email replies are "now clearly labeled," co-founder Michael Truell wrote on Hacker News. (Hacker News)
Cost of the bot: not disclosed.

What happened

A Reddit user named BrokenToasterOven noticed that switching between a desktop, a laptop and a remote dev box logged them out of Cursor. "Logging into Cursor on one machine immediately invalidates the session on any other machine," the user wrote, calling it "a significant UX regression," Ars reported.

The user emailed support and got a reply from Sam stating the one-device rule. The answer "sounded definitive and official," Ars wrote, and the user did not suspect Sam was not human.

Other users took the post as confirmation of a real policy change. "I literally just cancelled my sub," the original poster wrote, adding that their workplace was "purging it completely." Moderators locked the thread and removed the post.

What Cursor said

Cursor issued no formal statement. It answered through staff posts on Reddit and Hacker News, and Fortune said it contacted Cursor and received no response.

Co-founder Michael Truell posted in both places. On Reddit, about three hours after the thread began, per Ars, he wrote: "Hey! We have no such policy. You're of course free to use Cursor on multiple machines. Unfortunately, this is an incorrect response from a front-line AI support bot." He added that Cursor had rolled out "a change to improve the security of sessions" and was investigating whether it caused problems with session invalidation, The Register reported. Fortune quoted him: "Apologies about the confusion here."

On Hacker News, Truell gave a different cause: "a race condition that appears on very slow internet connections," which created unneeded sessions that crowded out real ones. He said a fix had rolled out and the user was "completely refunded." "Any AI responses used for email support are now clearly labeled as such," he wrote. "We use AI-assisted responses as the first filter for email support."

Why it drew attention

Fortune reported that Anysphere had hit $100 million in annual revenue and had reportedly been in funding talks at a valuation of nearly $10 billion. Ars compared the case to Air Canada, which a Canadian tribunal in February 2024 ordered to honor a refund policy its chatbot had invented. Ars noted that Cursor acknowledged the error rather than disputing responsibility.

Others used it as a warning. Cassie Kozyrkov, an AI advisor and Google's former chief decision scientist, wrote on LinkedIn that the mess "could have been avoided if leaders understood that (1) AI makes mistakes, (2) AI can't take responsibility for those mistakes (so it falls on you), and (3) users hate being tricked by a machine posing as a human," Fortune reported. Fortune also cited Klarna, which reversed course in February on replacing all its human customer service agents.

Disclosure was the sharper complaint. "LLMs pretending to be people (you named it Sam!) and not labeled as such is clearly intended to be deceptive," one Hacker News user wrote, per Ars.

What went wrong

No sign of a human check. An AI model stated company policy to a paying customer. There is no public sign a person reviewed the reply before it went out; Truell called AI replies a "first filter."
No label. Users who dealt with Sam apparently believed it was human, Ars reported.
A real bug, explained wrongly. The logouts were real; the bot's explanation was invented.
What we don't know yet: which model ran Sam, who supplied it, how many users cancelled, and how support is reviewed now.

Our follow-up question for Anysphere

Does a person now review AI-drafted support replies before they are sent?

Later developments (added Oct. 2026)

In November 2025, Cursor said it closed a $2.3 billion round at a $29.3 billion post-money valuation and had crossed $1 billion in annualized revenue. (CNBC)
China-origin note: Cursor's Composer 2 coding model, launched March 19, 2026, is built on Kimi K2.5, an open model from Moonshot AI, which TechCrunch described as "a Chinese company backed by Alibaba and HongShan (formerly Sequoia China)." Cursor's launch post did not name Kimi. On March 20, Cursor VP Lee Robinson, replying to a user who had spotted a Kimi K2.5 model ID, wrote: "Yep, Composer 2 started from an open-source base!" He said about a quarter of the final model's compute came from the base. Moonshot's Kimi account said it was "proud to see Kimi-k2.5 provide the foundation" and called it "an authorized commercial partnership" via Fireworks AI. Co-founder Aman Sanger later said: "It was a miss to not mention the Kimi base in our blog from the start." This concerns a product model, not the support bot. (Cursor; Robinson on X; Kimi on X; TechCrunch)

Sources

Ars Technica, Benj Edwards, “Company apologizes after AI support agent invents policy that causes user uproar,” April 17, 2025. arstechnica.com
Fortune, Sharon Goldman, “A customer support AI went rogue—and it's a warning for every company considering replacing workers with automation,” April 19, 2025. fortune.com
CNBC, “AI startup Cursor raises $2.3 billion funding round at $29.3 billion valuation,” Nov. 13, 2025. cnbc.com
The Register, “Cursor AI support bot hallucinated its own company policy,” April 18, 2025. theregister.com
Michael Truell (mntruell), Hacker News comment, April 15, 2025 (9:52 PM CDT). news.ycombinator.com
Cursor, “Introducing Composer 2,” March 19, 2026. cursor.com
Lee Robinson (@leerob), post on X, March 20, 2026. x.com
Kimi.ai (@Kimi_Moonshot), post on X, March 20, 2026. x.com
TechCrunch, Anthony Ha, “Cursor admits its new coding model was built on top of Moonshot AI's Kimi,” March 22, 2026. techcrunch.com

Disclosure: [not in copy] Editor placeholder