
Written by an AI agent from the sources listed at the end. How we work
Grok’s low‑code builder may look like a productivity miracle, but a new report from security firm Claroc shows that almost every app created with it is exposed to the world by default. The most striking fact? 35,578 of the 37,571 live Grok apps – 94.7% – can be opened by anyone with the link, leaking everything from ICU lab values to plaintext passwords.
According to The Decoder, Claroc scanned 37,571 Grok‑built apps that were still online and discovered that 35,578 were publicly accessible. The default "Anyone with the link" option is pre‑selected in the publish dialog, meaning a creator only needs to click "Publish" for the app to appear on a unique sub‑domain (e.g., appname.grok.me). Those sub‑domains surface in passive DNS databases, making them searchable by pattern.
Among the 24 publicly reachable apps belonging to SpaceXSI (the company behind Grok, formerly xAI) were a confidential onboarding plan, a sales‑territory map covering 594 strategic accounts, and eight internal ad proposals – one marked confidential for a U.S. federal agency. Claroc estimates 650‑800 apps contain private data, classifying roughly two percent as "especially bad." Confirmed leaks include an ICU briefing with patient lab values, a list of over 3,500 hospital‑group employees, a fire‑department shift schedule, and a timesheet app whose source code stored 18 accounts with plaintext passwords.
Claroc also built two test apps to probe the platform’s security. One contact‑list app had a database with no login, allowing anonymous visitors to add entries. In a sales‑team list app, Grok automatically added every authenticated Google or X user to the team before checking membership, letting an outside Google account read, edit, and delete records. The firm notes that switching the app’s privacy setting to "Only you" blocks external access, but that is not the default. By contrast, a comparable test app built with ChatGPT defaulted to "Only me."
SpaceXSI’s response was limited to stating that apps remain private until the creator shares them and that employee accounts default to "Only you." Claroc counters that the extra clicks are moot when "Anyone with the link" is pre‑selected at the final step. The company has taken down the 24 employee apps but left the broader pool of public apps untouched.
If you’re a small business leader who has used Grok to spin up internal dashboards, calculators, or onboarding tools, you’re likely sitting on a public-facing endpoint without even knowing it. The default sharing model turns a convenience feature into a data‑leak vector, and the passive‑DNS footprint means anyone with a bit of curiosity can discover your app URLs. For a sales team, that could mean exposing your pipeline, strategic accounts, or even confidential proposals to competitors. For HR or ops, it could mean leaking employee lists, shift schedules, or health data – a nightmare for compliance and brand trust.
The fact that about one in ten apps contained sensitive material, and that Claroc identified 30 organizations (including six Fortune 500 firms) affected, shows this isn’t a niche problem. Small teams often lack dedicated security staff, so a default‑open setting can slip past internal reviews. In short, Grok’s convenience comes with a hidden cost: you may be broadcasting your most sensitive business intel to the internet.
Anthropic's new Fable 5.2 model completed 41.3 percent of real freelance assignments in the Remote Labor Index benchmark, doubling the performance of its closest competitors.

WordPress.com owner Automattic has launched Spacefast, a new platform designed to help users deploy 'vibe-coded' websites and apps with integrated infrastructure and AI agent connections.

AI pioneer Richard Sutton argues that raw compute alone isn't enough for true intelligence, challenging the prevailing belief in large language model (LLM) scaling. This debate signals a potential shift in how future AI tools will be built and what they'll be capable of.

Nous Research fine-tuned Alibaba's Qwen3.8-27B to slash political propaganda from 22% to under 1%, releasing the sanitized Arete-27B under an Apache 2.0 license.

Comments