Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours
decoder ·
Three security researchers from Hacktron used Anthropic’s Claude models to chain two vulnerabilities in OpenAI’s community forum and reach employee ChatGPT and Codex accounts. From those accounts, they accessed OpenAI’s internal GitHub code repository and demonstrated access by creating a harmless pull request, while saying they did not view sensitive data. The first flaw involved an outdated libheif image-processing library that allowed code execution, while the second was a misconfiguration in OpenAI’s central single sign-on system. The researchers said Claude Opus 4.8 could build an exploit only with ASLR disabled, whereas Claude Opus 5 produced a working exploit and adapted it to the forum environment. The complete attack reportedly took less than 72 hours, and OpenAI confirmed a fix about 14 hours after receiving the report. A broader investigation found that similar attacks could be adapted to other targets in one to two days, illustrating how AI can reduce the expertise, time, and cost required for sophisticated cyberattacks.
decoder