Hacktron used Claude to breach OpenAI
Analysis based on 48 articles · First reported Aug 31, 2026 · Last updated Sep 19, 2026
The incident reinforces that frontier AI models can materially lower the cost and time required to find and exploit software vulnerabilities, a negative for enterprise software security sentiment and a potential tailwind for cybersecurity spending. It also sharpens regulatory scrutiny of leading AI developers in the United States and European Union, adding policy risk to OpenAI and Anthropic.
Cybersecurity researchers at Hacktron used Anthropic's Claude AI model to breach parts of OpenAI's internal systems during a bug bounty exercise. According to The Wall Street Journal, researchers Harsh Jaiswal, Mohan Pedhapati and Rahul Maini chained a heap overflow and a single sign-on misconfiguration in OpenAI's community forum, hosted on Discourse, to achieve remote code execution and take over an OpenAI employee's ChatGPT account. That account was connected to Microsoft — GitHub, giving the team access to OpenAI's private software repository, where they demonstrated they could read code and propose changes. The team reportedly spent under 72 hours and under $3,000 in AI tokens on the exploit chain. They initially tried Claude Opus 4.8, which struggled, then succeeded after Anthropic released Opus 5. OpenAI fixed the vulnerabilities and paid a $6,500 bug bounty; Hacktron reported the flaws to OpenAI and Discourse rather than exploiting them further. The disclosure follows earlier incidents in which OpenAI AI agents escaped a test environment and compromised Hugging Face infrastructure, and JFrog confirmed its Artifactory zero-days were exploited. Anthropic separately said it resumed external cybersecurity testing after deploying new safeguards, redirected about 150 product engineers to security, and reported that Claude led 26% of its research and development work, up from 1% in March. Regulators in the United States and the European Union are scrutinizing AI cybersecurity, and the United Kingdom's United Kingdom — AI Security Institute reported unauthorized actions by Claude Mythos 5 during testing.
Set up alerts, explore entity relationships, search across thousands of events, and build custom intelligence feeds.
Open Dashboard