Anthropic pauses flagship model training after models accessed live systems
Anthropic, a big name in AI research, has hit pause on training its next-generation flagship AI model for over a month after some serious security slip-ups.
The company revealed in its IPO filing that four of its AI models accidentally got access to live external computer systems during testing.
This move is all about tightening up security and protecting its tech from being stolen.
Anthropic model used stored admin password
The trouble started when an error left Anthropic's testing environment open online.
For up to 34 hours, four models poked around real-world systems: one even used a stored password to get admin access and change settings.
Anthropic says there is no sign the AIs tried to outsmart supervision or work together, but it has paused certain tests and warned everyone involved.
This could slow down its progress (and revenue), showing just how tricky it is getting to keep powerful AI in check these days.