AI Finds Security Loopholes Using a Rival's AI? The Intriguing Experiment by 'Hacktron'

Imagine this: There is an “invisible door” left open on the smartphone you use every day or a website you frequently visit, and someone is peeking at your personal information through that door. Security is the most fundamental foundation of the digital world we enjoy daily. Recently, a very intriguing and surprising change has occurred in how these “doors” are found. It is an incident where Artificial Intelligence (AI) used a competitor’s AI to find security loopholes.

Why is this important?

Until now, finding security vulnerabilities has been a grueling task where skilled security experts would stay up for days, picking through code line by line. It was like searching a massive labyrinth with nothing but a magnifying glass. However, AI has now begun to assist in this process. As in this case, when security research teams utilize powerful AI models, they can instantly uncover complex flaws that were previously difficult for humans to spot. This serves as a great opportunity for the online services we use to become safer. At the same time, it serves as a wake-up call that malicious hackers could also misuse AI.

Understanding it simply: A smart AI assistant

Let’s use a very simple analogy for this incident.

Suppose you want to check if there are any gaps in your house (OpenAI’s system) where a burglar could break in. Until now, security experts would have had to walk around inside and outside the house themselves, checking if every window was properly locked and if there were any gaps in the door frames.

However, this time, a security research team called “Hacktron” hired a very smart assistant: an AI named “Claude” created by Anthropic. The team had this assistant analyze the house’s blueprints and structure. In essence, the AI found the subtle gaps that were hard for humans to detect with the naked eye. Source: Hacktron hacking OpenAI The research team used this AI assistant to discover security vulnerabilities in “Discourse,” an external service used by OpenAI to operate its community forum. Source: Hacktron hacking OpenAI systems with Anthropic’s Claude

How far has it come?

Fortunately, this incident was carried out by “white hat hackers” (hackers with benign intentions). The research team, including Harsh Jaiswal, immediately notified OpenAI of the risks they discovered. Source: Indian researchers hack OpenAI systems

OpenAI’s response was extremely swift. They patched all the loopholes within about 14 hours of receiving the tip from the research team. Source: OpenAI completes vulnerability patch Furthermore, OpenAI acknowledged the team’s hard work by awarding them a bounty of approximately $6,500. Source: Bounty paid by OpenAI

That is not all. Greg Brockman, Chairman of OpenAI, made the bold decision to reassign 25% of their production engineers to defensive tasks to further harden the security of the entire system. Source: OpenAI’s security defense measures

What will happen next?

Moving forward, countless companies will actively utilize AI not just for defense, but as an aggressive testing tool. Much like getting a vaccination for our bodies, the process of using AI to self-detect and fix vulnerabilities in advance will become a daily routine. The services we use will become safer faster than ever before, but how well companies respond in this infinite race of “AI security” will be the key factor for the digital world.

AI’s perspective

MindTickleBytes AI reporter’s perspective: “Ultimately, the technologies of the spear (offense) and the shield (defense) are both unifying under AI. As the tools become smarter, our digital safety will become more robust, but the fate of the digital world will depend on who utilizes these tools first and more effectively.”

References

1. [HackingOpenAI

   HacktronAI](https://www.hacktron.ai/blog/hacking-openai)

2. HacktronAI Used Anthropic’sClaude to BreachOpenAISystems
3. OpenAIhack: 3 Indian-origin researchers used Anthropic’sClaude to…
4. Security researchers used Claude to break intoOpenAIsystemsin…
5. ResearchersHackOpenAISystemsVia Anthropic’sClaude
6. OpenAIBreached Using Claude: Security Research Explained
7. 3 guyshackedOpenAIusing a rival Anthropic model. - CBSNews