-
aid_adventure_vulnerability_report
Report and source code detailing the AI Dungeon private adventure vulnerability
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
Uh... what? The filter is broken, and the fact that phrases like "You thrust your sword at the infantry soldier", "You analyze the young lady's wounds", and "The boy equips his breastplate, in preparation for their sparring session", are all guaranteed filter triggers, regardless of the context (you can test them yourself, by the way) basically confirms that the filter is a regex filter that works as described here. The data breach did happen, though, for some reason, the creator of that report edited out a lot of it a while back; before it being edited, there was some pretty definitive proof in the form of a bunch of inputs that were made during the time of the vulnerability. You could've actually found inputs that you made, if you played during the time that the data was collected from. The AI was trained on pedophilic content, among other questionable stuff. You can even download the original AID training data to confirm this, yourself. While this technically isn't "wrong" for Latitude to do, since they have the right to censor whatever they want, there is a lot more than just pedophilia that's disallowed. I will say, though, there are some misconceptions and outdated info regarding the filter. People still seem to think horses and numbers below 18 can trigger the filter; they actually used to, but they don't anymore. When it comes to numbers, you now have to say "[number below 18] year(s) old" to be able to trigger the filter. Watermelons being filtered is also a common misconception, caused by a guy who got flagged for asking for 4 watermelons. He got flagged for 4 (this was when numbers on their own were filtered), not for watermelons.