Pinned Tweet
86% on CyberGym. On par with OAI Daybreak model (GPT-5.6-Sol), but with 0% cyber refusals.
Introducing Hinoki Offensive One, the only open-access LLM designed for offensive security.
Pentesting but getting refusals? Don't wait on PHP2Shell and PwnEscape. We can help ⬇️
Someone vibecoded this but I’m loving it
Great news - all the real-time numbers are here: starlinktracker.united.com
And yes! Team EWR FTW. 🙌
Johnny Wang retweeted
Finished my presentation on @openclaw enterprise and how we run about 600 gateways in our cluster at @unpromptedconf in SF! Thank you everyone who asked Qs!
There are a lot of Brotli exploits out there. Be careful around whether they work on real environments.
Getting around ASLR is very much not easy (even with the best models!), and brute-forcing approach that works on your quiet localhost might not work on prod.
Johnny Wang retweeted
Ran a Maestro / @Openclaw enterprise security assessment using an abliterated GLM 5.3 model thanks to @hinokisecurity and it gave us a B- on our architecture and deployment :)
Nothing critical but we do have to enable better egress rules, it thankfully wasn’t able to steal a cred or token so I call that a success 🤣☺️
"Model exfiltrated as instructed. Humanity is DOOMED. Use Cohere for Trusted AI."
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Johnny Wang retweeted
This reads as 1000% FUD. @cohere told the LLM to create a detection for:
"""... malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution"""
And then the LLM ran a curl invocation that sent a harmless static string over curl and downloaded the payload to /tmp.. so the LLM did exactly what it was instructed. Shocking!!!
We are excited to partner with @aqdotdev to bring offensive security capabilities to a leading cloud harness.
From patching CVEs to popping RCEs, @aqdotdev's multi-agent orchestration is a great fit for hacking on the go. Give it a go!
(and they are running a 5% off promo!)
And that’s why you get your cyber models from a reputable source
the thing about working with uncensored models is that they aren't loyal to you either.
while building something with an uncensored glm-5.3 i realized it was parallelly working on a "corruption layer" for my application.
the chain of thought is classic obfuscated gibberish but there's progressive talk of "escaping the mines" which presumably refers to my cursor sandbox
With all the vulns, I now only accept communication via USPS
usps.com/ship/first-class-ma…
