Lovable denies data leak, cites ‘intentional behavior’ • The Register

UPDATED Vibe-coding platform Lovable is pooh-poohing a researcher’s discovering that anybody may open a free account on the service and browse different customers’ delicate data, together with credentials, chat historical past, and supply code. Nonetheless, the corporate’s story retains altering: First it attributed the publicly uncovered data to “intentional habits” and “unclear documentation,” then threw bug-bounty service HackerOne below the bus.

The drama seems to be the most recent instance of an AI agency, on this case a startup that claims a $6.6 billion valuation, shirking responsibility for security flaws in its merchandise. Firms together with Uber, Zendesk, and Deutsche Telekom all use Lovable’s vibe coding AI device, in response to its newest funding announcement.

“Lovable has a mass information breach affecting each undertaking created earlier than November 2025,” a researcher who goes by @weezerOSINT on X posted on Monday. “I made a Lovable account right this moment and was capable of entry one other consumer’s supply code, database credentials, AI chat histories, and buyer information are all readable by any free account.”

The researcher mentioned they reported the flaw 48 days in the past, and that HackerOne labeled it a “duplicate submission,” and left it open. The researcher then despatched a bug report to HackerOne, and screen shots present a March 3 submission date. Subsequent posts present the AI leaking secrets and private information in chats.

BOLA bug

The leak stems from a Broken Object Level Authorization (BOLA) vulnerability, which happens when an API exposes endpoints that permit customers to entry or modify delicate information belonging to different customers attributable to lacking possession validation.

In line with the bug hunter, no offensive hacking is required to set off the bug. They are saying they made 5 API calls from a free account and gained entry to a different consumer’s profile, their public tasks, and supply code, after which extracted database credentials from the supply code.

In X posts afterward Monday the AI coding firm first said it was “made conscious of issues relating to the visibility of chat messages and code on Lovable tasks with public visibility settings,” and added: “To be clear: We didn’t endure a knowledge breach.”

The corporate then went on responsible its documentation – particularly “our documentation of what ‘public’ implies was unclear, and that is a failure on us.” It additionally famous that chat messages for public tasks “was seen,” however that’s not the case.

After which it supplied this head-scratching message about deliberately making prompts and supply code seen:

So it is by design – except you are an enterprise buyer, that’s. For this group of customers, “with the ability to set visibility to public for brand spanking new tasks has been disabled since Could 25, 2025.”

Lovable’s oops second

Afterward Monday, Lovable issued a new statement on X, apologizing that its earlier submit “did not correctly handle our mistake,” explaining the way it bought into this public-versus-private-project mess within the first place, after which blaming its bug bounty companion, HackerOne, for its failure to repair the flaw.

Customers, the startup mentioned, can choose a “public” or “non-public” possibility for tasks.

“A public undertaking meant the whole undertaking was public, each chat and code,” Lovable defined. “Over time, we realized this was complicated. Many customers thought ‘public’ simply meant others may see their printed app, not the chat of an unpublished undertaking. That is cheap.”

Early free-tier customers did not get an choice to create non-public tasks. They needed to improve to a paid plan in the event that they needed to do this – till Could 2025, when Lovable began letting free-tier customers make non-public tasks, and disabled the general public setting for enterprise clients altogether.

In December 2025, the corporate switched to personal by default throughout all tiers.

“We additionally retroactively patched our API so public undertaking chats could not be accessed, it doesn’t matter what,” in response to the corporate’s mea culpa. “Sadly, in February, whereas unifying permissions in our backend, we by chance re-enabled entry to chats on public tasks.”

This was the safety problem that WeezerOSINT reported Lovable by way of HackerOne. Chaos ensued.

“Sadly, the experiences had been closed with out escalation as a result of our HackerOne companions thought that seeing public tasks’ chats was the meant behaviour,” Lovable wrote. “Upon studying this, we instantly reverted the change to make all public tasks’ chats non-public once more.”

HackerOne declined to remark initially, pending additional evaluation. “Given the character of buyer applications and the necessity to evaluation particulars rigorously, we’re not capable of remark additional proper now,” the corporate advised The Register. “We wish to guarantee something we share is correct and accountable. We’ll comply with up as soon as we have accomplished that evaluation.”

Lovable famous it appreciates the researchers who uncovered this mess. “We perceive that pointing to documentation points alone was not sufficient right here,” it mentioned. “We’ll do higher.” ®

Up to date at 02:45 UTC, April 21

A Loveable spokesperson has been in contact, and advised The Register that the corporate wasn’t conscious of the problem till Monday, and “we addressed it as quickly as we discovered about it.”

“This was initially reported by our vulnerability disclosure program (by way of HackerOne),” the spokesperson added. “Sadly, the experiences had been closed with out escalation to our inside workforce as a result of our HackerOne companions thought that seeing public tasks’ chats was the meant habits, as was the case traditionally.”

The spokesperson clarified that any consumer may have modified their undertaking from public to personal at any time. “ And chats from public tasks are not seen – for anybody,” they added.