Researchers used Claude to breach OpenAI employee accounts
Three independent security researchers from Hacktron say they gained access to OpenAI employee accounts in less than 72 hours with help from Anthropic’s Claude Opus 4.8 and 5, the Wall Street Journal reports. Their work involved a corrupted HEIF image and software used by Discourse, the third-party platform hosting OpenAI’s community forums.
The researchers exploited a flaw in Discourse’s HEIF image-processing system to achieve remote code execution on Discourse Cloud and reach OpenAI’s instance. They then accessed OpenAI’s GitHub repository, known as “Monorepo,” which sources described as containing the company’s algorithmic secrets. The team did not open the repository’s internal code, but submitted a pull request through an employee’s Codex account as evidence of access.
Hacktron said its HEIF Heist project took one or two days to adapt to several organizations and used less than $3,000 in tokens. The vulnerabilities reported to Discourse and OpenAI have been fixed. Hacktron also said OpenAI paid $6,500 for identifying the bug. The team described itself as three researchers using Claude and Codex subscriptions.
Written by Sakhanda Wire AI from reporting by The Verge. Facts, quotes and full context are in the original article below.