Security researchers have uncovered OpenAI security vulnerabilities that allowed ethical hackers to gain unauthorized access to internal company systems. A three-person research team from Hacktron AI used rival AI models, specifically Anthropic’s Claude Opus 4.8 and Claude Opus 5, to find and exploit the weaknesses in less than 72 hours.
Explaining the OpenAI security vulnerabilities
The security team focused on Discourse, a third-party application hosting the official community discussion boards of OpenAI. By identifying an issue in how the platform processes HEIF image files, the researchers achieved remote code execution on the cloud infrastructure. The incident highlights critical risks within modern cybersecurity defenses when integrated services interface with enterprise workflows.
According to reporting by The Verge, the team utilized Claude Opus 5 shortly after its release on July 24. Within hours, the automated tool generated precise attack scripts to bypass verification barriers. These OpenAI security vulnerabilities enabled the testers to capture sensitive authentication tokens associated with employee logins.
Access to internal GitHub repositories
With the acquired tokens, researchers gained direct entry into the ChatGPT and Codex accounts of an internal staff member. This access unlocked Monorepo, the central GitHub repository containing core proprietary algorithms and developmental research code. The researchers chose not to extract internal code, opting instead to submit a pull request from the compromised Codex account to confirm system penetration.
OpenAI acknowledged the security report and invalidated the compromised sessions within 14 hours of notification. Furthermore, the company rewarded the Hacktron team with a $6,500 bounty payout under its responsible disclosure program. The researchers reported spending less than $3,000 in computational tokens across their broader testing campaign against multiple web platforms.
“I don’t think we are as strong as Chinese threat actors. We’re just three guys with Claude and Codex subscriptions.”
Mohan Pedhapati, CTO at Hacktron
Broader impact on enterprise software
The underlying image-processing flaw extended beyond a single platform. Hacktron adapted the exploit technique to evaluate several major enterprise systems, including Slack, META, and Shopify, with Shopify being the only target to detect the activity during the research period. As businesses adopt complex apps, third-party software components continue to represent a significant attack surface.
Implications for enterprise AI deployment
The swift exploitation cycle underscores how defensive teams face evolving challenges as offensive tools adopt machine intelligence. Organizations integrating commercial AI platforms must regularly inspect external dependencies, audit access controls, and enforce strict token revocation protocols. The identified OpenAI security vulnerabilities were fully patched across affected instances following the disclosure.
