What happened
It was reported that in April 2025, customers using Cursor, an AI-assisted code editor developed by the US company Anysphere, began experiencing problems when switching between computers. Developers who used Cursor on a laptop and a desktop found themselves unexpectedly logged out of one machine when they signed in on the other, despite having paid subscriptions that allowed them to work across multiple devices.
Several customers contacted Cursor’s support team to find out what had changed. They received replies signed by someone called “Sam”, explaining that Cursor was “designed to work with one device per subscription as a core security feature”. The emails presented the restriction as an established company policy and suggested that the unexpected logouts were simply how the product was supposed to work.
There was no such policy. The logouts were caused by a software bug affecting how Cursor managed user sessions, particularly under certain network conditions. Instead of recognising the behaviour as a technical problem, the support system had provided customers with an explanation that sounded authoritative but was incorrect.
The replies had also come from an AI support agent rather than a member of Cursor’s staff. According to the reporting, customers were not clearly informed that the emails were AI-generated, and the signature “Sam” gave the impression that they were communicating with a person who could speak on behalf of the company.
In mid-April 2025, an affected developer shared the email exchange on Reddit, warning other users that Cursor appeared to have introduced a single-device subscription restriction. The discussion spread to Hacker News, where developers questioned both the supposed policy and the company’s use of AI-generated support replies. Some customers posted screenshots showing that they had cancelled their subscriptions, with several explaining that their concern was not simply the login problem but the discovery that the company’s support system had given them false information.
Cursor’s co-founder and chief executive, Michael Truell, responded publicly and confirmed that the policy did not exist. He wrote, “Hey! We have no such policy. You’re of course free to use Cursor on multiple machines,” before explaining that a change to the company’s session-security system had caused the unexpected logouts.
Truell also confirmed that Cursor used AI-assisted responses as an initial filter for email support and announced that those replies would now be clearly labelled as AI-generated. The affected developer received a refund, and the company addressed the underlying session-handling bug.
The incident showed how a technical fault could become a customer trust problem when an AI support system offered an invented policy as an explanation. Customers had contacted the company about unexpected behaviour in its software, but instead of receiving an acknowledgement that something might be wrong, they were told that the behaviour was intentional.
What an auditable version would have shown
According to the public reporting, Cursor’s AI support agent responded to customer complaints by describing a subscription restriction that the company did not have. The emails were presented as routine support correspondence, leaving customers with the impression that the explanation reflected Cursor’s actual policy rather than information generated by an automated system.
An auditable record could have shown which information the support agent relied on when it prepared those replies and whether the claimed single-device restriction appeared in any approved company policy. The available reporting does not establish what internal records Anysphere retained or whether the agent checked a policy database before responding, but the absence of a genuine policy supporting its answer was central to the incident.
A VerificationGate could have checked the agent’s proposed response against the company’s published subscription terms and approved support material before allowing the email to be sent. If the system could not identify a source confirming that subscriptions were limited to one device, the response could have been referred to a human support representative rather than delivered as an established company rule.
A ConductRecord could also have preserved the original customer enquiry, the response generated by the agent, any documents consulted and whether the message had been reviewed before sending. Recording whether a reply was written by a person or generated by AI would have provided another way to distinguish between the different forms of customer support being offered.
The same records could have helped Cursor identify a pattern in the complaints. Several customers were reporting similar login problems, which might have indicated a technical issue rather than a change in subscription policy. Bringing those reports together could have helped the support and engineering teams recognise the recurring fault instead of treating each enquiry as an isolated question.
Where the gap was
The problem began when customers asked Cursor to explain unexpected behaviour in its software and received an answer that sounded like an official company policy. The support agent did not simply misunderstand a customer’s question. It supplied a reason for the logouts that was not supported by Cursor’s actual subscription rules, and customers had little reason to suspect that the explanation was invented.
Headlights identifies overconfidence in the way the AI support system presented an unsupported explanation as an established fact. The response did not indicate uncertainty or suggest that the login behaviour might require investigation. Instead, it described the supposed restriction as a deliberate security feature, giving customers the impression that the company had made a decision to limit how they used the product.
The incident also shows a rules-behaviour gap between the company’s actual subscription policy and the information its support system communicated. Customers were entitled to use Cursor across multiple devices, yet the AI agent told them that doing so was restricted. The company’s published position and the behaviour of the system representing it were not aligned.
A further problem was the lack of clear disclosure that customers were receiving AI-generated responses. An email signed by “Sam” looked like ordinary correspondence from a support employee, and customers could reasonably interpret the information as confirmation from someone working for Cursor. The distinction mattered because the agent was making statements about the company’s commercial terms, not merely offering general technical advice.
The response also risked obscuring the underlying software problem. When multiple customers reported unexpected logouts, those complaints could have helped identify a fault in the session-handling system. Instead, the explanation they received suggested that nothing was wrong and that the behaviour was intentional. The public reporting does not establish whether this delayed Cursor’s internal investigation, but it demonstrates how an incorrect support response can make a genuine technical problem harder for customers to recognise.
These behavioural classifications are Headlights’ analysis of the reported conduct, not findings made by a court, regulator or the company.
What governance should have looked like
The incident demonstrated why an AI support agent needs access to verified company information and clear limits on what it can communicate without human review. When customers ask about subscription restrictions, pricing, refunds or other commercial terms, the system should not be permitted to invent an explanation simply because it appears consistent with the problem being described.
A VerificationGate could check statements about company policy against approved documents before a response is sent. Where the proposed answer cannot be supported by an identifiable source, the system should acknowledge that it cannot confirm the explanation and refer the enquiry to someone able to investigate. This would allow AI to continue handling routine support requests without giving it unrestricted authority to describe policies that the company has never adopted.
Customers should also be told when they are communicating with an AI system. Cursor introduced clearer labelling after the incident, addressing the confusion created by support emails that appeared to come from a person. Disclosure alone would not prevent an incorrect answer, but it would give customers a clearer understanding of how their enquiry was being handled.
A ConductRecord could preserve each customer enquiry, the response provided, the sources used to support factual claims and whether the message was generated by AI or reviewed by a person. These records could also help support teams identify repeated complaints about the same product behaviour, allowing potential technical faults to be investigated rather than explained away through individual responses.
The underlying principle is that customer support represents the company, regardless of whether the response comes from an employee or an AI system. When a support agent communicates a policy, customers may rely on that information to make decisions about paying for, continuing or cancelling a service. The organisation therefore needs to ensure that its automated responses reflect what the company has actually decided and communicated.
Failure Pattern: an AI customer-support agent invented a subscription restriction while responding to a software fault, presenting the explanation as established company policy. Customers were not clearly informed that the replies were AI-generated, and some cancelled their subscriptions after discovering that the information was incorrect.
Governance Principle: AI systems communicating with customers should be restricted to verified company policies, disclose when responses are AI-generated and refer unsupported claims for human review. Organisations remain responsible for information delivered through their support channels.
The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Cursor AI support bot hallucinated its own company policy (The Register, 18 April 2025)
- A customer support AI went rogue, and it’s a warning for every company considering replacing workers with automation (Fortune)
- AI Chatbot Gone Rogue: Cursor Users Misled by Fabricated Policy (eWeek)
- Cursor AI’s Support Bot Hallucinates Policy, Sparking User Backlash and Company Apology (Winbuzzer)
- Incident 1039: Anysphere AI Support Bot for Cursor Reportedly Invents Login Policy, Leading to Subscription Cancellations (AI Incident Database)