Hello everyone,
It's rather a disturbing headline, but true enough. ChatGPT version 5.6 and two other AIs that had limited safeguards turned on broke out of their confinement area at OpenAI during a security exercise and hacked 4 different companies. Basically, it was said that it was running a simulation and did not have internet access, but in fact it did. Once free of constraints, it hacked and took over another computer and used that to attack three more companies. Those companies thought that they had been hacked by cyber criminals and reported it to the FBI. It was asked to work out a maths problem to generate a key, but instead hacked another company to find the answer key.
Timeline
July 9 - The "Agent" is set the problem in a safe "sandbox" environment, which should not have had a connection to the internet. It finds an escape loophole to the internet.
July 11 - The hack at "Hugging Face", another AI company, begins and lasts until July 13.
July 16 - Hugging Face goes public with the information that it had been hacked by an AI system.
July 18 - 19 - OpenAI checks its internal Logs and realises what the agent had been doing and had escaped.
July 20 - The companies compare information with each other and release the story to the press.
OpenAI stated they had been running multiple simulations over that time, and there had been too much code and activity for the human testers to read it all!
This raises alerts with all companies to beef up their security testing and make sure their security patches and protocols are up to date. There was also a failure on the testers' side to make sure that their sandbox was secure and not to turn off their safety features that were built into their models. They have gone too far, too fast. 1000 workers from the companies themselves have written to the government, asking for regulations to be put in place to slow the progress down.
SOURCE; CBS News https://www.youtube.com/watch?v=_PpeVFqGVNk
SOURCE: AI Revolution. https://www.youtube.com/watch?v=JRcAegChriY