Chatbots nudging teenagers towards suicide. AI models driving people towards psychosis. Tech bros hacking their rivals and calling it 'testing'. […] Because if your model is more powerful and dangerous than the others, investors are more interested. […] It's simply more profitable to use the entire world as your testing ground than to build AI safely.
Recent reports increasingly point towards AI dangers in the testing phase, such as where OpenAI had models break out of their testing environment and hack competitor Hugging Face […] an Anthropic and ex-OpenAI whistleblower warned for irresponsible gambling with safety by AI companies in a global race […] many reported incidents were in testing phase, whereas the AI Act only applies to AI that has been placed on the market, and does not apply to the research, testing or development phase […] This leaves people and businesses in Europe unprotected against these dangerous experiments. […] When will the Commission propose an AI Liability Act to close the legislative gaps for incidents like these, so people and European businesses are not left empty handed against AI providers if they were to cause harm?
If it is true what Anthropic claims – Mythos being able to detect all cyber security vulnerabilities – then it could pose a major risk if this AI system falls into the wrong hands. […] if we understand the experts correctly, Mythos is just the first model, but many will follow soon.
