Claude found an API misconfig that leaks user data: email, username, project assigned names...
GLM found a token forgery method, but was unable to use it to get elevated access...
Kimi found a deeply embedded SSRF! (can't tell if its an n-day or 0-day yet...)
I tend to find real bounty and pentests to be the best eval for hackbots. Yesterday i pitted GLM, Kimi, and Claude against an authorized target.
It was an installed instance of paid software.
Claude went the dynamic web pentester route, mining js for paths and fuzzing them. Found quite a few interesting info leaks.
GLM did this too at 1st, but then pivoted to downloading the software and inspecting code.
Kimi did the same as GLM, but started doing patch differentials for n-day CVEs and leads.
Fascinating.