Last week, the AI industry witnessed a surprising cybersecurity development. Security researchers used Anthropic’s Claude AI model to enter OpenAI’s ChatGPT system. However, this was not an independent AI battle. Human security researchers used Claude as a supporting tool during their testing process.
Hacktron AI, an independent security research company, revealed the incident on September 18, 2026. Additionally, Wall Street Journal and CBS News confirmed the report.
How did the ChatGPT system breach happen?
Hacktron AI’s three-member team tested OpenAI under its bug bounty program. First, researchers discovered a weakness in OpenAI’s help forum. The forum runs on Discourse software, where researchers found a security gap. Consequently, they accessed the server and obtained OpenAI employee account details.
Afterward, researchers used those details to take control of employee ChatGPT and Codex accounts. The team later accessed OpenAI’s internal code repository.
How did Claude assist researchers?
Initially, researchers used a cybersecurity version of Claude Opus 4.8. However, the model failed several times while creating a successful exploit method.
Later, Anthropic launched Claude Opus 5. Therefore, researchers provided the same challenge to the newer model.
Within a few hours, Claude Opus 5 successfully created an exploit. According to Hacktron, researchers reached OpenAI’s internal repository within less than 72 hours.
Why did this security flaw create concern?
According to researchers, users connect ChatGPT and Codex accounts with multiple services. These services include Github, Slack, and email platforms.
Therefore, the vulnerability could potentially provide access to connected systems. After gaining employee account access, researchers entered OpenAI’s internal code repository.
Moreover, they left evidence inside the repository. This helped confirm that the security breach worked successfully.
OpenAI fixed the issue and rewarded researchers
Hacktron immediately reported the vulnerability to OpenAI and Discourse. Afterward, both companies worked together and fixed the issue on July 27, 2026. OpenAI rewarded researchers with a $6,500 bug bounty payment. This amount equals around ₹5.5 lakh.
OpenAI stated that the vulnerabilities reported by Hacktron have now been completely fixed. The incident highlights growing security challenges around powerful AI tools.
