Censys Finds AI-Built Student Apps Collecting Sensitive Data Without Basic Safeguards
Summary
A Censys scan-based investigation identified AI-built education apps collecting student data, including homework photos, essays, voice, and screen activity. Some lacked privacy policies; one exposed default demo credentials and shared a host with tunneling services.
Key points
- Censys identified a population of AI-built websites marketed to students and found that more than 75% were deployed within the past year.
- About 30% showed evidence of direct data intake, such as file uploads, camera or microphone access, or connections to backend storage.
- Examples included apps soliciting homework photos, essays, GPA details, voice recordings, and screen sharing; some did not link a privacy policy.
- The investigation found no conventional malware, phishing, or other malicious indicators across the broader population.
- One language-learning platform displayed default admin credentials, admin / 123qwe, and shared its host with NPS tunneling services, a DERP relay, and more than 30 open ports.
- Censys cautions that the exposed student-data systems create security and privacy risks even without evidence of malicious activity.
Article Details
- Attack Vectors
- Lingua AI's login page displays the default demo credentials admin / 123qwe.
- Lingua AI's host exposes more than 30 ports, including NPS tunneling services on ports 8080 and 10080; the article describes this combination as atypical, but reports no threat labels or malware indicators.
- AI-built student applications collect sensitive information through file uploads, camera and microphone access, screen sharing, and password-based accounts.
- Defensive Notes
- Review exposed services and port ranges on systems that handle student data; the article identifies a dense, multi-service host as an unusually broad attack surface.
- Change default credentials and review access controls for education applications that expose login pages.
- Assess privacy and data-handling practices for applications collecting students' files, images, voice, video, or personal information; the article notes that some examples lacked linked privacy policies.
Products
BubbleThe remaining sites are built on Express, Bubble, Next.js, and traditional stacks where the AI label fired on other signals.CaddyThe backend is legit: Kestrel (.NET) across five high-numbered ports, an OpenAI API proxy on port 3000, Caddy reverse-proxying traffic on the front.ExpressThe remaining sites are built on Express, Bubble, Next.js, and traditional stacks where the AI label fired on other signals.Firebasecalls for camera or microphone access, and references to backend storage services like Supabase or Firebase.KestrelThe backend is legit: Kestrel (.NET) across five high-numbered ports, an OpenAI API proxy on port 3000, Caddy reverse-proxying traffic on the front.LovableIn 2026, a new generation of AI development tools like Lovable and Replit performs a remarkably similar trick.Next.jsThe remaining sites are built on Express, Bubble, Next.js, and traditional stacks where the AI label fired on other signals.ReplitIn 2026, a new generation of AI development tools like Lovable and Replit performs a remarkably similar trick.SupabasegetUserMedia calls for camera or microphone access, and references to backend storage services like Supabase or Firebase.