AI hacks AI? How security researchers breached OpenAI using Claude

Imagine you are the security chief guarding a massive castle. You know there is a tiny crack in the castle wall, but it would take a person days of sleepless nights to find it. What if a smart assistant suddenly appeared and said, “I can find that crack and open the door for you in just three days”?

In the world of artificial intelligence (AI), this recently happened for real. Three researchers from the Indian security startup ‘Hacktron AI’ successfully hacked OpenAI’s internal systems by leveraging Anthropic’s AI model, ‘Claude’ [Source 5, Source 9].

Why does this matter?

More important than the news that “someone was hacked” is the fact that the ‘subject’ of the hacking has completely changed. Until now, finding vulnerabilities in complex systems required highly skilled hackers working in teams and investing a significant amount of time.

But now, AI is taking over that role. In this incident, the researchers succeeded in accessing OpenAI’s internal code system and private GitHub repositories in less than 72 hours [Source 1, Source 6]. Even more shockingly, this entire hack cost less than $3,000 [Source 9]. This suggests that AI has drastically lowered the barrier for cyberattacks, meaning security threats are now much closer to us all.

Easy to understand: The digital detective, AI

Why is AI helpful for hacking? Simply put, AI acts as an excellent ‘digital detective.’

Equipped with the ‘Transformer’ technology (an AI architecture that identifies relationships between words in sentences and analyzes vast amounts of data) that we use commonly, AI reads and analyzes tens of thousands of lines of code and security documents in an instant. It is like reading a hundred thick books in just one minute and finding a tiny contradiction or logical hole within them.

In this attack, the researchers targeted a vulnerability in a third-party forum plugin called ‘Discourse’ [Source 1, Source 7]. Initially, they used ‘Claude Opus 4.8,’ but the attack failed. However, when they switched to Anthropic’s newly released ‘Claude Opus 5’ model, it broke through the security barrier that the previous model could not resolve and created functional exploit code [Source 4, Source 6]. As AI gets smarter, the success rate of attacks is increasing alongside it.

Current Status: Ethical hacking

Of course, this hack was an ‘ethical hack’ performed by security researchers with good intentions. They proved the vulnerabilities in OpenAI’s system and received a bug bounty [Source 3, Source 11].

However, what we must remember is that not only Claude but also OpenAI’s own model, ‘GPT-5.6 Sol,’ was employed in this attack [Source 3]. In other words, hackers are pitting AI models against each other or combining them to design more powerful attacks. This demonstrates the ‘duality’ of AI. AI can be a solid shield for security, but at the same time, it can be a powerful weapon for attack.

What happens next?

It is highly likely that an ‘AI vs. AI’ security war will unfold in the future. Attackers will try to use AI to find security vulnerabilities faster, while defenders will have to build even smarter AI to build defensive walls in real-time.

From a user’s perspective, we must be more vigilant about security when using AI. While companies strengthen AI security, it has become even more important for individuals to follow basic security rules, such as not carelessly clicking on suspicious links or files. This incident reminds us that security in the AI era is not just a program issue, but has become a daily risk management task for all of us.

MindTickleBytes AI Reporter’s View

This incident proves that AI is evolving from a simple tool into an ‘agent’ that autonomously explores system vulnerabilities. While there was a positive experiment by security researchers, the repercussions are difficult to predict if a malicious attacker gets their hands on this technology. Moving forward, as much as AI development is crucial, a technical leap in ‘defensive AI’ that can block security threats caused by AI is urgently needed.

References

1. OpenAIhackedbyresearchersusingAnthropic’sClaude

   LinkedIn: https://www.linkedin.com/news/story/openai-hacked-by-researchers-using-anthropics-claude-8638745/

2. ResearchersusedClaudetohackOpenAIemployees’ ChatGPT…: https://www.theregister.com/security/2026/09/18/researchers-used-claude-to-hack-openai-employees-chatgpt-accounts/5297517
3. OpenAI‘ethicallyhacked’ with help of Anthropic’sClaudechatbot: https://www.theguardian.com/technology/2026/sep/18/openai-hacked-anthropic-claude-chatbot
4. ResearchersusedAnthropic’sClaudetohackintoOpenAI: https://techcrunch.com/2026/09/18/researchers-used-anthropics-claude-to-hack-into-openai/

5. ClaudewasusedtohackOpenAI

   Generative AI Newsletter: https://newsletter.genai.works/p/claude-was-used-to-hack-openai

6. Security researchers used Anthropic’s Claude to hack OpenAI’s …: https://the-decoder.com/security-researchers-used-anthropics-claude-to-hack-openais-internal-systems-in-under-72-hours/
7. Security researchers used Claude to help them hack into OpenAI: https://www.theverge.com/ai-artificial-intelligence/997444/openai-hack-claude-heif-heist
8. Hackers Used Anthropic’s Claude to Break Into OpenAI: https://www.wsj.com/tech/ai/hackers-used-anthropics-claude-to-break-into-openai-b40ba883
9. Three Indian researchers used Claude to hack into OpenAI in …: https://thetechportal.com/2026/09/18/openai-hacked-using-claude-hacktron-ai-indian-security-researchers/
10. AI security experts say theyusedClaudetohackChatGPT - CBSNews: https://www.cbsnews.com/news/claude-hack-chatgpt-anthropic-openai/
11. Indian-originresearchersusedClaudeAItohack…: https://www.livemint.com/technology/tech-news/indianorigin-researchers-used-claude-ai-to-hack-openai-s-systems-got-6-27-lakh-bounty-11789753784566.html