A team of three independent security researchers at the cybersecurity firm Hacktron demonstrated how advanced artificial intelligence can accelerate cyberattacks by gaining access to internal OpenAI systems in under 72 hours. Using Anthropic’s Claude Opus AI models, the researchers discovered software vulnerabilities and breached employee accounts, spending less than $3,000 in AI computing tokens to perform the work.
The breach began at Discourse, a third-party service hosting OpenAI's community forums. Hacktron exploited a vulnerability in how the platform processes HEIF image files. Within 14 hours of Anthropic releasing Claude Opus 5 on July 24th, the team achieved remote code execution—a critical flaw allowing attackers to run unauthorized commands on a remote system—and took control of OpenAI’s forum cloud instance.
From Discourse, the researchers accessed OpenAI employee accounts and reached the company’s GitHub repository, known as "Monorepo," which contains OpenAI’s core algorithmic secrets. Although the team chose not to view internal code inside Monorepo, they confirmed their level of access by sending a pull request to the repository using an employee’s Codex account.
Hacktron's research project, dubbed "HEIF Heist," adapted the exploit against multiple major platforms in just a few days, but only one target company, Shopify, detected the attack. Hacktron reported the security flaws to Discourse and OpenAI, both of which have patched the vulnerabilities. OpenAI paid Hacktron a $6,500 bug bounty, though Hacktron CTO Mohan Pedhapati warned that if three independent researchers with off-the-shelf AI subscriptions could succeed, sophisticated foreign threat actors pose an even greater danger.