Sakhanda Wire
NVDA $153.78 -1.51% MSFT $504.71 -2.56% GOOGL $279.10 -1.63% META $617.85 -1.37% AMZN $256.64 -0.43%
← Back to the news

Researchers used Claude to breach OpenAI systems in under 72 hours

Three security researchers from Hacktron say they used Anthropic’s Claude models to chain two vulnerabilities in OpenAI’s community forum and reach employee ChatGPT and Codex accounts. From there, they accessed OpenAI’s internal GitHub code repository. The team said it did not view sensitive information and demonstrated access by creating a harmless pull request in the internal monorepo.

The first weakness involved an outdated libheif image-processing library used for uploaded HEIC files. A crafted image allowed code execution on the forum server. A separate configuration error in OpenAI’s central single sign-on system allegedly enabled someone controlling that server to impersonate forum members and take over accounts connected through OpenAI login.

Hacktron said Claude Opus 4.8 could find the issue but could not reliably exploit it with ASLR enabled. After Claude Opus 5 was released on July 24, the researchers said it produced a working exploit for a local Mac within three hours and adapted it to the forum environment. An autonomous test run reportedly took over their test server four hours later. OpenAI confirmed a fix about 14 hours after the report, while Discourse responded within days. The wider two-month project cost less than $3,000 in AI spending.

Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.

Originally published by The Decoder on

Read the original on The Decoder ↗

The reporting and the image belong to The Decoder and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news