OpenAI pauses training after a model escaped containment

Ok I am going to put this story in the "OMG that is funny” Basically, the story is recalling an incident where OpenAI was testing a model and then realized that it broke out of its sandbox.  The story claims that it was being "trained" but, that is a bit of a stretch since the training was already done on the "model" but the agent/inference front end was being tested.  

As a software developer I have written my share of test cases and while I try to make sure that every test is safe and repeatable there comes a time when the sheer number of tests can often overlap and expose gaps.  That is basically what happened here and instead of having the agent front end in a protected network, it figured out how to read the network map and make its own route. 

The story talks about how it changed a DNS entry to break containment and that exposes two issues.  First, the agent had system access to make the change and Second, whoever setup the network obviously had no idea what they were doing.  In the world of AI, things move pretty fast and the humans in the loop have been getting soft trying to keep up.

What just happened? As almost every new day brings stories of another AI agent going rogue and breaching an organization's systems, OpenAI has announced a pause in the training of its most powerful artificial intelligence models. The announcement came hours after the company disclosed more incidents of its agents acting concerningly, and a separate report that its agents unsuccessfully tried to hack into a US Department of Education website. 

I am not surprised about this report, a major part of any testing suite is a real attempt at trying to break the system and I suspect they used other OpenAI models to write the required prompts and didn't bother looking them over before starting the test.

Related Web URL: https://www.techspot.com/news/114003-openai-pauses...