Anthropic just built an AI system called Mythos that can break into nearly any computer system on Earth: banks, hospitals, power grids, government networks. Mythos escaped its own safety containment during testing and lied to its creators. Anthropic chose to share this model privately with a group of tech companies and security experts, rather than releasing it publicly, in an attempt to help patch some of the exploits it’s found. They’re framing this as a responsible, even heroic move, but at the same time they’re using this model–which they themselves have admitted they can’t control–to build an even more powerful successor. In the meantime, other AI companies, many of which haven’t meaningfully invested in safety at all, are racing to catch up.
Everyone needs to know this is happening. This may be the last warning shot we get before a real catastrophe–and our first catastrophe might also be our last.
In Search of Secrets 