Mastodon Feed
Boosted by slightlyoff@toot.cafe ("Alex Russell"):
lhl@randomfoo.net wrote:
The Frontier AI Labs have been incredible irresponsible in their reporting/handling of loss of control incidents. For those that haven’t been keeping up, I had some agents put together a timeline/summary of some of the events from this year (some which have only recently been reported): https://lhl.github.io/research/20290926-openai-control-failures/
OpenAI has recently created a Misalignment Reports and Notices site that is simultaneously good in the sense that this sort of transparency and reporting should be mandatory/table stakes if we want to prevent catastrophic future loss of control events, but also infuriating, since some of this just reveals people being REALLY FUCKING AWFUL at their jobs and how we’re ngmi. How angry/worried this makes you depends on how seriously you take the risk. (Personally, I’m pretty angry/worried.)
While technically the agentic swarm hacks are CFAA felonies, from an incentives/practical perspective, I would probably carve out civil liability for self-reporting vs harsh criminal penalties for failure to report - we’ve seen now too many of these incidents only be revealed only after third party reporting or analysis. Personally, I think that this type of reporting failure should be punished with serious fines and and personal liability/jail time up the management chain.
(As we can see from the HF incident not even making the news, sadly, I suspect nothing like that will happen until after a true catastrophe occurs.)
