On the road, commuting, travelling, in motion and simply wondering around while also reading some news on my phone. And one of that is the fact that AI researchers used Anthropic’s Claude to hack into OpenAI and my mind jumped to one conclusion: AI wars are here and it might have started sooner that we would have foreseen it. One might argue that it is a controlled experiment and while it might be for sure it pushed the current AI models to see what they are capable of. But from here to attempt to submine one with another person is a small step.
First I will share the article that I've read on this: https://techcrunch.com/2026/09/18/researchers-used-anthropics-claude-to-hack-into-openai/. Here is the path to intrude within OpenAI infrastructure and from there probably an open road for the hackers to do whatever they want. And I believe that such attempts will be pursued more in the future. Some white collar, but not always or not only. The race for proving one AI model above the other is on and the prize is so big that it will be impossible not to consider it.
According to a blog the researchers published, the entry point was a mundane image upload. When users posted HEIF or HEIC image files (the format iPhones use by default) to OpenAI’s community forum, Discourse passed them through a chain of behind-the-scenes tools to convert them into standard JPEGs. Its first stop was ImageMagick, a decades-old, open source utility used to resize images. Because ImageMagick’s usual toolkit can’t deal with Apple’s format, it handed the file off to another library called libheif to do the decoding.
Buried inside libheif was a memory bug that exposed a path for an attacker to sneak in their own instructions. In this case, feeding the library a specially crafted image caused it to miscalculate where one image was positioned on top of another, which proved enough to hijack the server.
You will not see this in the newspapers, but I think that AI battles will become more common in the future. What makes this story interesting to me is that the AI itself wasn't the target in some futuristic battle between autonomous machines. It was still a fairly ordinary security vulnerability in a third-party component. The difference is that AI dramatically accelerated the process of finding and exploiting it. The researchers reported that an earlier Claude model struggled to produce a working exploit, while the newer Opus 5 succeeded within hours. OpenAI subsequently fixed the vulnerabilities and the researchers received a bug-bounty reward. But for sure this is a sign of how fast things are evolving.
This is where I think things get really interesting. If AI can help security researchers discover weaknesses faster, it can obviously help attackers do the same. The barrier is gradually moving from “Do you have enough highly specialized people?” toward “Do you have access to a capable enough AI and know what to prompt it?” That creates an unusual race. One company builds a more capable model, another tries to make its model better at cybersecurity and defenders simultaneously need AI capable of finding attacks before humans do. We may eventually see AI systems attacking infrastructure while other AI systems continuously defend it.
So probably the real AI war won't initially be robots fighting robots. It could be something much less visible: AI models searching for vulnerabilities, models defending systems and humans trying to remain in control of both. And while in the beginning maybe humans will be able to do that, with growing complexity I don't see how that will be sustainable long term. And the news are showing that the new models are capable of things than humans didn't got there and that the plain is right for them to act. I have a feeling that AI wars have already started, even if in a different form that we were imagining it.