Cybersecurity startup Hacktron AI utilized Anthropic's AI, Claude, through OpenAI's official security testing program to safely verify internal OpenAI systems.
Imagine this: someone secretly breaks into the work messenger or internal company bulletin board you use every day. However, instead of a malicious hacker, this intruder is a “white hat hacker” (a security expert hired to legally identify and report corporate security vulnerabilities). This is exactly the kind of intriguing event that recently occurred in the AI industry.
Researchers from cybersecurity startup ‘Hacktron AI’ successfully breached OpenAI’s security systems using Claude, the AI chatbot from competitor Anthropic.
Why does this matter?
This event signifies that AI has become the most powerful “weapon” and “shield” in the realm of hacking and security. While past hacking relied entirely on human intuition and effort, AI’s vast knowledge and rapid reasoning capabilities are completely transforming security testing methods. It is particularly significant that AI has begun to play a key role as an assistant in verifying the safety of the AI services we use and determining how far internal information can be exposed.
In simple terms: A hacker with a capable AI assistant
Let’s use an analogy for this incident. Suppose there is a detective tasked with investigating a massive fortress (OpenAI’s security system). To understand the fortress’s structure, the detective hires a very intelligent and articulate “assistant” (Claude).
| The assistant helped the detective find paths into the fortress and quickly read complex internal documents (such as corporate bulletin boards) to find important clues. The Hacktron AI research team found OpenAI’s internal security vulnerabilities with the help of this AI assistant, Claude. “Ethical hacking” refers to activities like this, where vulnerabilities are not exploited but are instead politely reported to the company to help prevent future incidents. [Source: OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot | OpenAI | The Guardian](https://www.theguardian.com/technology/2026/sep/18/openai-hacked-anthropic-claude-chatbot) Source: AI security experts say they used Claude to hack ChatGPT - CBS News |
What did they verify?
The Hacktron AI research team performed this work as participants in OpenAI’s official security program. This program is a bug bounty system that rewards security researchers for discovering gaps in their systems. Source: Cybersecurity Researchers Hack Into OpenAI Using Anthropic’s Claude Chatbot - SSBCrack News
| With Claude’s help, the research team successfully accessed the ChatGPT accounts of some employees and even entered ‘Discourse,’ the platform where OpenAI employees held internal discussions. [Source: Hacktron AI Researchers Use Anthropic’s Claude To Hack OpenAI, Access ChatGPT Account: how 19 outlets framed it | NewsCord](https://newscord.org/article/hacktron-ai-researchers-use-anthropics-claude-acc–Story_20260918_ResearchersusedClaud2b04bbd6) Source: AI security experts say they used Claude to hack ChatGPT - CBS News |
| Through this process, the team was able to identify key data regarding where OpenAI’s source code (the blueprint for their computer programs) is stored and managed. They also sent harmless test requests (Pull Requests) to OpenAI’s GitHub (a source code sharing service) to see how the system handled them. However, the researchers clearly stated that they did not actually download any source code, and that the entire process was for the purpose of testing the system’s security. [Source: OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot | OpenAI | The Guardian](https://www.theguardian.com/technology/2026/sep/18/openai-hacked-anthropic-claude-chatbot) Source: AI security experts say they used Claude to hack ChatGPT - CBS News |
What’s next?
This case demonstrates that AI can accelerate human security work by tens or hundreds of times. The security market will likely see an even more intense battle of wits between “hackers using AI” and “security teams using AI for defense.” Leading companies like OpenAI are expected to continue operating these ethical hacking programs to further harden their AI systems. From our perspective as users, as these “safety checks” become more frequent, we can hopefully look forward to AI services becoming gradually more secure.
References
-
OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot OpenAI The Guardian (https://www.theguardian.com/technology/2026/sep/18/openai-hacked-anthropic-claude-chatbot) -
OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot Business News finwire.io (https://finwire.io/news/business-news/openai-ethically-hacked-with-help-of-anthropics-claude-chatbot) -
Hacktron AI Researchers Use Anthropic’s Claude To Hack OpenAI, Access ChatGPT Account: how 19 outlets framed it NewsCord (https://newscord.org/article/hacktron-ai-researchers-use-anthropics-claude-chat-account–Story_20260918_ResearchersusedClaud2b04bbd6) - OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot - The Bold News (https://theboldnews.com/openai-ethically-hacked-with-help-of-anthropics-claude-chatbot/)
- AI security experts say they used Claude to hack ChatGPT - CBS News (https://www.cbsnews.com/news/claude-hack-chatgpt-anthropic-openai/)
- Cybersecurity Researchers Hack Into OpenAI Using Anthropic’s Claude Chatbot - SSBCrack News (https://news.ssbcrack.com/cybersecurity-researchers-hack-into-openai-using-anthropics-claude-chatbot/)
- To destroy the system
- To safely test OpenAI's security vulnerabilities
- To leak company secrets
- ChatGPT
- Claude
- Gemini
- Personal photos of employees
- Source code
- Advertising data