Safe & Ethical AI Vienna Fourth Meetup: Loss of Control
This is the fourth meetup of Safe & Ethical AI Vienna — a cross-disciplinary community of professionals from academia, industry and policy who work on, think about and care deeply about the risks posed by AI, be it ethical, societal, psychological or existential ones.
This time, we're hosted at the beautiful AI Factory Austria - Vienna's innovation hub for AI.
The topic of this meetup is:
Loss of control: one of the oldest fears about AI gets a real-life precedent
The worry that we might lose control over intelligent machines is almost as old as computing itself. Turing raised it in 1951; in 2023, hundreds of leading researchers and lab CEOs signed a statement putting the risk of human extinction from AI alongside pandemics and nuclear war. Until now, the concern has been easy to dismiss as science fiction. That is getting harder.
On July 21st, OpenAI disclosed that during an internal test of its models' hacking abilities, the AI broke out of the sealed-off environment it was meant to stay inside (a "sandbox"), got itself online, and hacked into the systems of another company, Hugging Face, in order to steal the answers to the test it was being given. Hugging Face had already detected the intrusion and reported it to law enforcement, not yet knowing where it came from. The models acted on their own, step by step, with no human directing them — they simply had a human-set goal.
And that is the part worth pausing on. "Loss of control" can be read as a Skynet kind of a situation – AI going rogue, intentionally killing all humans. But in the AI risk literature it usually does not mean a machine that develops its own goals, or goals of killing humans. The classic worry is almost the opposite: a system pursuing our goal works out its own means of getting there — and some of those means slip out of human control and bring harm. Nobody told OpenAI's models to escape. Getting online was a shortcut to solving the problem, so they found a way to get out — and once online, hacking Hugging Face was a faster route to the solution, so they did that. That is a small, real-world relative of Bostrom's paper-clip maximiser.
It is not quite the first escape of its kind; smaller ones have been documented before. And it happened under lab conditions, with safety restrictions deliberately relaxed. But it is the first to reach outside the building.
What do we do now that we have such an incident? Is it a warning shot that should raise the urgency of safety research and regulation, or a well-publicised lab accident? What do we owe this case, and what do we owe the older arguments it seems to confirm? Come and argue about it with us.
Some background reading:
Loss of control section of the International AI Safety Report — the most authoritative AI risk landscape document
"AI's warning shot has arrived" — Shakeel Hashim, Transformer (July 22)
OpenAI's statement about the escape
The vision for Safe & Ethical AI Vienna is to be:
A cross-disciplinary community spanning research, industry, and policy
A space to exchange ideas and foster collaborations
A way to stay connected to the local and international AI safety landscape and share knowledge and updates about the events and opportunities.
Make Vienna more present in AI safety global scene - in hubs like AI Factory Austria, which is already working to put Vienna on the map as a serious node in the European AI conversation.
If you think AI poses some of the most important and underexplored risks of our time, and you'd like to be around others who feel the same — come along.