Soooo, Anthropic developed this new frontier model, Claude Mythos Preview, that found so many vulnerabilities while it was being tested that they decided not to release it.
They believe that AI models have reached a “level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities”. Not concerning at all </sarcasm>.
Instead of releasing this model out into the wild, Anthropic has decided to continue testing it and they created an initiative called Project Glasswing.
They invited some of the largest tech companies to join them in this initiative to test the model(s), identify existing vulnerabilities, and, hopefully, carve a path forward.
Some of the companies that have already joined are -
They believe that what they’ve already found could reshape cybersecurity. They haven’t said in which direction, though…
They already found vulnerabilities in every major operating system. Also, in every major web browser.
There is quite a bit more to this but I will keep it short in the name of Quick Tips.
If you are interested in learning more, here is the project’s introduction post from Anthropic - https://www.anthropic.com/glasswing
By the way, the project is named after the glasswinged butterfly (greta oto) that is able to hide in plain sight.
Here is a pic -