Security researchers used Claude to help them hack into OpenAI
Three researchers used Anthropic's Claude to hack OpenAI employee accounts in under 72 hours via a third-party forum flaw.Summary
A team of three independent security researchers from Hacktron reported using Anthropic's Claude Opus AI models to gain access to OpenAI employee accounts. The researchers exploited a vulnerability in the HEIF image processing system of Discourse, the platform hosting OpenAI's community forums. Within less than 72 hours of Claude Opus 5's launch, they achieved remote code execution on Discourse Cloud and accessed OpenAI's instance. They demonstrated their access by sending a pull request from an employee's Codex account to OpenAI's internal "Monorepo" GitHub repository, though they did not access its code. The "HEIF Heist" project was adapted to target multiple other companies, including Slack and Meta, for under $3,000 in tokens, and only Shopify detected the attempt. The vulnerabilities were subsequently fixed, and OpenAI paid Hacktron a $6,500 bug bounty. The researchers emphasized they are not a major threat actor, highlighting the potential power of AI-assisted hacking.
(Source：The Verge)