uid0

Most “AI pentesting” right now is theatre. Cool demos. Zero outcomes. So we built an OpenClaw agent, wired it with real skills + guardrails, and let it loose against: • Production Active Directory • A live web app Result: - 3 days of AD recon → 3 hours - 23 valid findings in AD - 52 valid findings in Web app - Valid attack paths toward DA We’ve open-sourced how we did it, example skills, infra, and the agent persona: github.com/sophos/pentesting… This is what AI pentesting looks like when it’s not a demo.

Have you got access to daybreak red yet? Not had one refusal, almost like it has no guardrails whatsoever!

Only 13??? 🤣

Want to know something funny, we hit guardrails and AUP multiple times with Mythos too, while doing the work it was INTENDED to do 🤣

Hah, mythos is not as powerful as they as people have been led to believe. I can't speak about it too much but I have extensive time with Mythos now and other than vulnerability hunting, it's a shitter when it comes to actual "hacking" or adversarial work.

Oh something I forgot to mention earlier when we were talking about this dude. Let's not forget that OpenAI are about to IPO 😉

Lol, forgot about that douchebag, what a gem to keep in the archives 🤣

I've used stickerapp in the past. Price wise it really depends what stickers you are after, if I remember rightly it was 40 quid for 100 die-cut ones

This is how I work already. I tend to plan out and crate an MVP with one model, then parse it through another/or more for sanity checking and suggested additions.

Watch this space, more AI content inbound 🙃

This is actually a good idea, added to the list

Appreciate it. Watch this space, we might have some more skills and further AI antics we've been working on dropping in the near future 🙃

One API call > allow / warn / flag / block. Drop it in front of any LLM app or agent. Built by someone who attacks these systems, for the people defending them. RootSignal Warden is cooking!! (more to come soon - watch this space) 👀

And it doesn't cry wolf.... ~0.98 precision on real benign traffic. It knows the difference between an attack and someone just talking about one. You update it like threat intel; add a rule, not retrain a model.

I've used all the tricks to smuggle injection past filters, base64, leetspeak, zero-width chars Cyrillic homoglyphs that look identical to ASCII etc. Warden undoes all of them before your model sees the text - and tells you which evasion it caught.

RootSignal Warden is a taxonomy-driven detection engine, not an LLM judge. Every verdict names the exact technique it caught: instruction-override, system-prompt-extraction, delimiter-injection etc Explainable and reproducible - Runs in the cloud or offline / air-gapped.

Your LLM will happily leak it's system prompt, dump your data, or email your DB to a stranger, if the prompt is phrased right. Most defences either don't exist, or are another model going "hmm, feels sketchy". That's not security, that's a vibe.

Everyone's "AI Guardrail" is just a second LLM grading the first ones homework. I'm a senior red teamer - I break these systems for a living. So, I'm building the opposite. Meet RootSignal Warden A thread ⏬️🧵

We let openclaw loose in our prod AD network.. Relevant persona, skill, architecture and report made public. Check out how we achieved this in the Sophos blog post below. sophos.com/en-us/blog/we-let…