Independent security researchers from Hacktron AI used Anthropic's Claude Opus 5 to breach OpenAI's defences, according to a report by The Wall Street Journal. The three-person team exploited two critical vulnerabilities as part of an OpenAI bug-bounty program.
The researchers gained access to multiple OpenAI employee ChatGPT accounts and subsequently entered the company's software, including a GitHub organisation linked to an employee's Codex account. The initial entry point was a flaw in Discourse, the third-party software powering OpenAI's community forum, specifically related to HEIF/HEIC image file processing.
Hacktron AI reported their findings to OpenAI, which has since resolved the issues and awarded the startup $6,500. The researchers noted that an earlier version, Claude Opus 4.8, struggled to create a working exploit, but Opus 5 succeeded within hours of its release.
Matt Fredrikson, CEO of Gray Swan, commented that such tools, available for around $200 a month, can enable hacking into companies like OpenAI. Hacktron founder Mohan Pedhapati stated that AI is reducing the expertise needed for exploit development, potentially shortening work that once took months to days.