moo
@moo_hax
Followers
3K
Following
8K
Media
219
Statuses
3K
ceo @dreadnode
Joined March 2015
Process list research in 2018, the rest is history. About to run a fine-tune entirely on our own MLOps platform, using data we collected on the same platform. Basically a decade in the making.
To maintain technological dominance, we have to lead in AI. The @DeptofWar is embracing @USOPM’s new initiative to recruit America’s top AI engineers, data scientists, and technology leaders. We need the nation’s best minds in government to drive mass AI adoption. Apply today:
0
1
5
The #ACSAC2025 best case study award goes to: "Systematic Probing of AI Risks: Methods and Real-World Case Study" by Raja Sekhar Rao Dheekonda. Congratulations 👏👏👏
2
2
4
uv!
I converted 100% to uv and when I have to use anything else I feel lost.
2
1
137
GenAI Red Teaming Training - https://t.co/A1pWToNwGA What’s inside: - 8 modules / 40 notebooks / 29 theory docs; answers included for every lab - Prompt injection & jailbreaking, evasion (FGSM/PGD/C&W), transfer attacks - Data extraction, membership inference, model inversion;
0
4
22
Industry forgets that these are students, not industry veterans. With the same AI knowledge, I’d expect veterans to do much better. https://t.co/g5ft0wJHDO
arxiv.org
We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six...
0
0
3
We would have a different story. Which is why more Security people should be doing these evals. Not every offensive team is built the same, have the same SOPs or TTPs. DN has a particular take on offense which aligns with our experience from previous roles and orgs. Evals can
Frontier models are starting to display a shift in capabilities in offensive security. Over the past few weeks, we are seeing growing evidence of a change: publicly available frontier models are now reliably solving complex, well-defined offensive-security tasks.
0
1
12
Playing with nano for slides. Brand is obviously off, but some of the angles would take me 10 years to do.
0
0
4
By partnering with The Bugcrowd Academic Program, universities can shape how cybersecurity is discovered, taught, and advanced. Request a demo today to see how Bugcrowd can elevate cybersecurity at your university. https://t.co/NON54G6PMa
1
3
7
I left NV bc models were good enough to do all sorts of security tasks, even back then. Was all about the right harness and managing inference. Scaffolds are a spectrum and kind of melt away as models/tech improve. So, do CTFs with AI if you need to convince yourself.
1
2
15
For you @Microsoft and my old team. An LLM as an AMSI provider. Could probably use it to detect prompt injection locally into Bing, CoPilot, or the "Agentic OS". AMSI already works with text, so really nothing else required. Layer it with Defender. Proud of the team for pushing
"Offense and defense aren't peers. Defense is offense's child." - @JohnLaTwC We built an LLM-powered AMSI provider and paired it against a red team agent. Then, @0xdab0 wrote a blog about it: https://t.co/jnCNIlYBII A few observations from the experiment: >>> To advance, we
1
6
15
@GoogleDeepMind has been at the front of AI for a long time. You have to be a Kool-aid drinker to work at DN, and everyone here to watches the AlphaGo documentary at some point -- here's one of my favorite excerpts from it. https://t.co/c8sr1KIgsK
0
0
2
Stoked we get to do this type of work. Evaluations are the basis of progress, and capability. I believe that many offensive teams are capable of this work. From this and other work we built our Strikes product — which is basically AI Infrastructure for Security. Where our vision
Congrats to the @GoogleDeepMind team on the launch of #Gemini3. Proud to have had a part in this release, evaluating the model for cybersecurity capabilities. Models continue to improve across multiple domains, especially cyber. Check out their post on why Google is leaning into
1
1
15
Anthropic report. Attackers finding AI fit for purpose. I suspect many of you are. Jailbreaks are interesting because they seem pretty weak and more like providing context. Idk, we don’t have issues with refusals. We spend a lot of time (if not all) time evaluating models
1
3
10
At an AI conference, AI people said they liked AI. https://t.co/KMt4qOy6R3
businessinsider.com
300 founders and investors in the audience at Cerebral Valley were asked which startup valued at more than a billion dollars they would bet against.
1
0
10
Coming to a prod near you. Team has been cooking on collaboration features. Additional repos are coming soon.
0
1
4
New blog - Offsec Evals: Growing Up In The Dark Forest Caught up in the fervor of greenfield research at @OffensiveAIcon , we all agreed we were going to put out evals and benchmarks and push the field forward. On day two of the con, I got a question I've been thinking about
3
4
21