Security researchers from Hacktron AI successfully breached an OpenAI employee's ChatGPT account using the Anthropic Claude model. The attack began on July 23, when the team identified a vulnerability in the Discourse platform, which hosts OpenAI's community forums. By utilizing Claude Opus 5, the researchers generated exploit code for the CVE-2026-45788 vulnerability.
The exploit allowed the team to obtain session tokens that were valid for both ChatGPT and OpenAI's GitHub services. Through these tokens, the researchers gained access to 'Monorepo,' a repository containing internal software documentation and algorithms. The team stopped the attack after realizing they had accessed sensitive data, having only submitted a proof-of-concept pull request.
OpenAI confirmed the incident and stated that both the Discourse and internal vulnerabilities have been resolved. The company has restricted the scope of community login tokens and revoked affected sessions. Following the report, OpenAI paid the researchers a bounty of $6,500. The company also announced that 25% of its production engineers have been reassigned to security tasks.
The incident highlights the growing role of AI in cybersecurity. Hacktron AI CTO Mohan Pedhapati noted that the attack was executed using standard subscriptions to Claude and Codex. Meanwhile, industry reports suggest that illicit access to similar AI-powered hacking tools can be purchased on underground forums for as little as $800.
