People who went skeptical on “AI-orchestrated APT” report by Anthropic most likely haven’t tried to use their recent models coupled with custom MCP tools for intrusion security purposes. As person who did tried — I can say yes, they’re really that good. Getting Sonnet to hack shit for you isn’t even a challenge anymore, while gaining the same performance level from self-hosted models still is

@d_olex sounds interesting but implausible, want to record a video of you doing this and put it somewhere?

@whitequark No

@d_olex classic AI user response

@whitequark You have to sign NDA 😇

@d_olex okay fair

@d_olex @whitequark

@d_olex what kind of attack are we talking? Stealing credentials, logging in, and then dropping backdoors?

@siguza Recon, profiling, lateral movement and stuff like that — almost everything except initial access inside target perimeter (haven’t tried to automate this part yet since at this point I want to get rid of SaaS models rather than make it even more advanced)

@d_olex okay, so what I'd consider "anything but the interesting part" 😅

@siguza Yes, there’s no magic in there. The goal is to automate boring routine stuff that even mediocre red team operator can do, not to jump above one's head

@siguza But gaining perimeter access with password spraying or public exploits is also mediocre-level task, so…

@d_olex @whitequark Honest question: how is the NDA compatible with SaaS models?

@buherator @whitequark It’s mostly related to “custom tools” part, I haven’t experimented much with public ones because even most advanced models can’t do much if your tooling suck for the first place

@buherator @whitequark But they are capable of spotting and fixing impacket bugs on the fly tho 😂

@d_olex @whitequark oh I think I misunderstood! I'm concerned about how NDAs the Red Team is usually subject to (the target, data from their systems, etc) can be compatible with these third party services.

@buherator @whitequark I plan to get rid of any third-party services before this will go beyond the test lab