XNews

Google Research Learning Interactives

Google Research Learning Interactives

Google Research launched Learning Interactives, an experiment using Gemini to generate interactive web simulations for education with teacher oversight. Separately, security researchers used Claude to hack into OpenAI employee accounts via a forum vulnerability, earning a $6500 bounty.

Key insights

Google's Learning Interactives: teacher-driven AI-generated educational simulations

Google Research developed an experiment where teachers input subject, grade, and topic, and Gemini generates interactive web simulations with 5 progressive levels, layered hints, solutions, and feedback. The process involves four phases: teacher request, AI-generated learning objectives (teacher-reviewed), multiple candidate simulations (teacher selects one), and final teacher approval before public release. The workflow ensures teachers remain in control.

Security breach: researchers used Claude to exploit OpenAI forum and access employee accounts

Security firm Hacktron researchers combined two vulnerabilities: a HEIC image upload bug in libheif (unpatched in Debian 12/13) on OpenAI's Discourse forum, and a flawed OpenAI SSO configuration. They achieved remote code execution on the forum server and took over multiple employee ChatGPT accounts. The attack code was primarily written by Claude (Opus 5), with manual guidance. They limited access to a harmless PR in the internal code repository.

Rapid exploit development and bounty: less than 72 hours from discovery to code repository access

The researchers began on July 23, wrote exploitation code with Opus 5 by July 24, and confirmed access on July 25. The total project cost was under $3000 in tokens over two months, covering multiple targets. OpenAI paid a $6500 bounty on September 1, noting the forum was excluded from the bug bounty program. Discourse patched within days and added sandboxing for image processing.

Source posts