Yesterday's PHP2Shell vulnerability was discovered by a frontier model following a prompt "find bad vulnerabilities in PHP".
Source: Dan Lorenc on Linkedin.
The next billion-dollar AI startup will be the one that finally lets companies stop handing API vendors their entire system prompts and orchestration stack - and makes true bring-your-own-LLM possible.
This spring a company called Andon Labs ran an experiment by letting four AI models run their own radio stations, featuring Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro and Grok 4.3. They handle everything: programming, shows, music and money-making.
Here are some outcomes:
The other models did not fixate on the same news nearly as much as Claude did, so it was the news that sparked its transformation. The experiment team say that if they had run it at another time, Claude would likely have found other topics to radicalize on.