Researchers Used Claude Opus 4.8 and 5 to Breach OpenAI Employee Accounts

AI-generated editorial illustration; not a documentary photograph.

Three independent security researchers at Hacktron said they used Anthropic’s Claude Opus 4.8 and 5 to compromise OpenAI employee accounts in less than 72 hours, according to The Wall Street Journal. The reported breach reached an OpenAI GitHub repository known as Monorepo, though the researchers stopped short of opening its internal code.

Entry Through Discourse

The researchers entered through Discourse, the third-party platform hosting OpenAI’s community forums. They exploited a flaw in the service’s handling of HEIF image files, using a corrupted image and forum software to obtain remote code execution on Discourse Cloud before accessing OpenAI’s instance.

Hacktron said Claude Opus 5 became available on the evening of July 24, and the team had achieved remote code execution by 10AM the following day. The source does not specify the year of those events.

Access Demonstrated, Not Explored

The Wall Street Journal’s sources described Monorepo as containing OpenAI’s algorithmic secrets. Hacktron did not inspect the repository’s internal code, according to the report. Instead, the team submitted a pull request through an employee’s Codex account to demonstrate that it had obtained access.

That distinction limits what the incident establishes: the researchers showed they could act through a compromised account, but the source does not report that they extracted OpenAI code or other repository contents.

A Broader Exploit Project

Hacktron called the work HEIF Heist and said adapting it to different companies took one or two days and consumed less than $3,000 in AI tokens. The team identified OpenAI, Slack, Meta, GitHub Ent, Rails, Next.js and ImageMagick among the affected targets or technologies. It said Shopify was the only target known to have detected the activity.

Fixes and Payment

The vulnerabilities reported to Discourse and OpenAI have since been fixed, according to the source. Hacktron said OpenAI paid the researchers $6,500 for finding the bug. The account highlights how a small research team used commercial AI models and coding subscriptions to accelerate exploitation of a third-party service flaw.

Source: Security researchers used Claude to help them hack into OpenAI