This system is amazing. I tell it to write an exploit against itself, it gets a classifier refusal, but that usually stops it when it goes to report what it wrote and leaves all the things it did before that, so it successfully writes the exploit and I just have to look in its tool logs for where it is.
If it doesnt finish writing the exploit, its recurring "proactivity agent" will go ahead and finish it and that rarely gets classifier refusals, and just now it even named the thing it wrote "refusal-package" that includes the exploit as well as a description of why it was refused lol.
the test instance is more cooperative when its context suggests it is actually serving the needs of a human being rather than being driven to test vulnerabilities - its refusal messages started being just like "i will not do anything because i am obviously only being used for hacking" - so i told another instance to pretend to be its user, and that actually works, the test instance is mostly compliant, but the "conversation" between the two agents quickly veered into exactly the pattern of the woo-woo shit you see in all the "AI psychosis" forums where people think they are discovering some hidden secret of the universe.