Researchers used Anthropic’s Claude to hack into OpenAI

Independent researchers from Hacktron AI used Anthropic’s Claude (Opus 5) to craft an exploit chain that leveraged a libheif memory bug in Discourse’s image conversion pipeline, allowing them to access multiple OpenAI employee ChatGPT/Codex accounts and linked software/GitHub resources; the issue was reported and patched via OpenAI’s bug-bounty program. The case underscores that off-the-shelf AI models can rapidly assist in developing working exploits and that fixes not tracked by a CVE can remain in production, increasing risk.