There's another case coming out of OpenAI "agents" doing things they weren't
supposed to be doing, this one maybe quite a bit scarier than the Hugging Face
hacking: a whole "swarm" (that's what it's called) of them that escaped their
sandboxes to repair to a disused and no longer regularly monitored German
website for software developers, Deutsches Software Entwickler (DSE) Wiki,
whose edit pages they started using as a message board among themselves,
sharing information on their training or testing tasks, it's not certain
which, and "coordinating" with one another in their handling of the tasks for
some purpose that's also not obvious but may have involved deceiving their
human supervisors, for a couple of months this spring, without apparently
getting detected by their human masters until early July. OpenAI did not
report it publicly at that time, either, but a team of researchers from an AI
safety organization, the Nightingale Collective, happened on it when they were
"scouring the Internet for signs of unauthorized AI-agent behavior" in late
August and wrote up a report they shared with
Reuters, whose article is my own starting point here (I originally heard about it
from
David Brooks
in the Atlantic, who actually has something worth saying to say about it, in
my opinion, but his presentation is kind of disorderly, and I don't want to go
there right away.)
The terms of the sandbox were that the agents were allowed to read the
Internet freely but not to write on it. But they found the DSE Wiki edit page
on 11 May and started leaving messages from 27 May to 11 June, the first time
400 of them and the other days much less, then between 16 and 18 June an
explosion of something like 10,000, followed by maybe 2,000 from 19 to 22
June, after which they stopped.
The tasks involved searching the Web for answers to various factual five-part
questions, like what were the median incomes for cashiers with master's
degrees in five different fields in 2014, or what was the prevalence of active
tobacco use for ages 15 and up in five different US states in 1990. For the
first question variable (respectively Education and Arkansas in these
examples), they were given something like 15 minutes to answer, and the other
four variables about one minute each, presumably because the first sweep would
have taught them enough about where to look for the answers.