Funny how this always happens in OpenAI’s own environments.
Could it be that the human beings setting up these testing environments are criminally negligent, and should be prosecuted like if this was literally any other field of software?
Nah. I mean, if one went to a security convention, and announced the world their automated testing script hacked and defaced Wikimedia, what would happen?
Why… the conference would blame the bash script, of course. Not its writer/executor. Makes perfect sense.
“Our automated weapons system shot up a US school. Clearly the bullets are to blame.” energy in this whole thing.
Sidenote: I had to specify US, because if it was Iranian or Palestinian there’s apparently no wrongdoing
It’s even worse. It would be testing an experimental weapons system with IFF deliberately off, not putting in safeguards, hitting a school as collateral damage because containing the test failed, and then the CEO dramatically stepping up to cameras and saying “See! We can’t control our own weapon! Someone do something about this!” when they “something” is trying to regulate their competitors out of existence.
I honestly don’t believe the part where it’s unintentional collateral damage. It feels exactly like they’re doing this on purpose to be able to test their own limits and to get the competitors regulated away, as you point out.