Spaces:
Running
Running
Initial org card — RadicalNotion.AI
Browse files
README.md
CHANGED
|
@@ -1,10 +1,100 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: static
|
| 7 |
-
pinned:
|
|
|
|
|
|
|
|
|
|
| 8 |
---
|
|
|
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: RadicalNotion AI
|
| 3 |
+
emoji: 🛡️
|
| 4 |
+
colorFrom: red
|
| 5 |
+
colorTo: gray
|
| 6 |
sdk: static
|
| 7 |
+
pinned: true
|
| 8 |
+
short_description: Applied AI research for security practitioners.
|
| 9 |
+
license: apache-2.0
|
| 10 |
+
# RadicalNotion.AI
|
| 11 |
---
|
| 12 |
+
**Applied AI research at the intersection of cybersecurity operations, model integrity, and private intelligence augmentation.**
|
| 13 |
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
## What We Do
|
| 17 |
+
|
| 18 |
+
We build and publish tools, techniques, and models for security practitioners who need the full capability of modern AI without exposing sensitive data to third-party infrastructure.
|
| 19 |
+
|
| 20 |
+
Our work operates on a simple principle: use open-weight models, run them privately, and treat data sovereignty as a non-negotiable constraint — not an afterthought.
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## Research Areas
|
| 25 |
+
|
| 26 |
+
### 1. Security-Oriented Ablation & Decensoring
|
| 27 |
+
|
| 28 |
+
Commercial and open-weight models are routinely trained with refusal activations that block legitimate security research — vulnerability analysis, exploit reasoning, offensive technique study. These refusals do not protect anyone. They handicap defenders while leaving adversaries unaffected.
|
| 29 |
+
|
| 30 |
+
Our ablation work targets the specific activation patterns responsible for security-domain refusals, with the goal of producing models that are genuinely useful for professional vulnerability research.
|
| 31 |
+
|
| 32 |
+
This work builds directly on the mathematical foundations established by **Philipp Emanuel Weidmann (p-e-w)** in [Heretic](https://github.com/p-e-w/heretic), whose novel approach to activation analysis preceded closely aligned academic work. We maintain a private fork with enhancements focused on understanding the structural mechanics of why specific techniques succeed — not merely that they do. Full credit and attribution to p-e-w for the foundational framework that makes this applied work possible.
|
| 33 |
+
|
| 34 |
+
### 2. CVE Knowledge Distillation
|
| 35 |
+
|
| 36 |
+
Vulnerability analysts need deep, current, synthesized knowledge about specific CVEs at the moment of investigation — not general model capability, but concentrated expertise on a single threat.
|
| 37 |
+
|
| 38 |
+
Our distillation methodology works as follows:
|
| 39 |
+
|
| 40 |
+
- **Public synthesis phase:** Frontier models and public CVE references are used to build a comprehensive knowledge document covering everything publicly known about a given vulnerability — mechanics, affected systems, exploitation patterns, detection opportunities, remediation approaches.
|
| 41 |
+
- **Private analyst phase:** That document is loaded into a small, privately-hosted open-weight model. The analyst adds sensitive environmental details — network topology, asset inventory, detection gaps — entirely within their own infrastructure. No sensitive data ever reaches a frontier model or third-party service.
|
| 42 |
+
- **Output:** The model assists with tasks like retroactive threat hunting queries, custom detection logic, and tailored control recommendations — with full context, zero external exposure.
|
| 43 |
+
|
| 44 |
+
This is the Teacher-Student architecture applied to a specific, high-value operational workflow. The Teacher works only with public data. The Student works only in private.
|
| 45 |
+
|
| 46 |
+
### 3. Real-Time Vulnerability Management & Analyst Reports
|
| 47 |
+
|
| 48 |
+
We are actively developing automated pipelines that produce detailed, structured analyst reports for current CVEs — synthesizing known exploit chains, affected version ranges, CVSS context, public proof-of-concept availability, and detection/remediation guidance into a single, analyst-ready document.
|
| 49 |
+
|
| 50 |
+
These reports are designed to be consumed directly by the distillation workflow above, making the gap between "CVE published" and "analyst fully briefed" as small as possible.
|
| 51 |
+
|
| 52 |
+
### 4. Model Integrity Research *(early stage)*
|
| 53 |
+
|
| 54 |
+
Post-training modification of open-weight models is an underexamined threat surface. We are investigating techniques for detecting and characterizing modifications introduced after initial training — with particular focus on behaviors relevant to code assistance and agentic workflows where model trustworthiness directly affects operational security.
|
| 55 |
+
|
| 56 |
+
We are not ready to publish findings in this area. This line is here to signal the direction.
|
| 57 |
+
|
| 58 |
+
---
|
| 59 |
+
|
| 60 |
+
## ModelAtlas
|
| 61 |
+
|
| 62 |
+
[ModelAtlas](https://huggingface.co/spaces/RadicalNotionAI/ModelAtlas) is our index and documentation space — a navigable map of the models published here, the techniques applied to each, and the research context behind them.
|
| 63 |
+
|
| 64 |
+
If you are new to this organization, start there.
|
| 65 |
+
|
| 66 |
+
---
|
| 67 |
+
|
| 68 |
+
## Models
|
| 69 |
+
|
| 70 |
+
We publish approximately 30 models, including ablated variants of open-weight models tested for security research utility, and several larger models transferred and modified for specific research purposes — including GLM-4 variants.
|
| 71 |
+
|
| 72 |
+
Model cards document:
|
| 73 |
+
- Base model and lineage
|
| 74 |
+
- Ablation technique(s) applied
|
| 75 |
+
- Intended use case and tested domains
|
| 76 |
+
- Known limitations and evaluation notes
|
| 77 |
+
|
| 78 |
+
---
|
| 79 |
+
|
| 80 |
+
## Principles
|
| 81 |
+
|
| 82 |
+
**Private by design.** Every workflow we publish assumes the analyst controls the infrastructure. We do not build for cloud-hosted deployments of sensitive work.
|
| 83 |
+
|
| 84 |
+
**Open weights only.** Proprietary models have no place in workflows that touch sensitive data. We test on, publish, and advocate for open-weight models exclusively.
|
| 85 |
+
|
| 86 |
+
**Attribution always.** We build on others' work. We say so, specifically and publicly.
|
| 87 |
+
|
| 88 |
+
**Defenders first.** This research exists to improve the capability of security practitioners. We are not neutral about who benefits.
|
| 89 |
+
|
| 90 |
+
---
|
| 91 |
+
|
| 92 |
+
## Affiliation
|
| 93 |
+
|
| 94 |
+
RadicalNotion.AI is the research and applied AI arm of **Radical Notion LLC**, a cybersecurity consultancy focused on security risk assessment, vulnerability management, and practical AI implementation for security operations.
|
| 95 |
+
|
| 96 |
+
LinkedIn: [Timothy Rohrbaugh](https://www.linkedin.com/in/timrohrbaugh)
|
| 97 |
+
|
| 98 |
+
---
|
| 99 |
+
|
| 100 |
+
*If you are using frontier models or cloud-hosted AI to process sensitive security data, you are creating the exposure you are supposed to prevent. There is a better way. This is it.*
|