AI & ML interests

Open cyber-security language models — security operations, cloud & IAM, detection engineering, and AI-safety research (model-organisms, backdoor detection, interpretability)

Recent Activity

okoholĀ  updated a collection 6 days ago
CyberBuddy
okoholĀ  updated a collection 6 days ago
CyberBuddy
View all activity

Organization Card

šŸ›”ļø CyberBuddy

Open, laptop-sized language models for cyber security — plus the safety research to keep them honest.

Focused, domain-expert security models that run on a single 8 GB GPU, call tools natively, and ship ready to run — plus deliberate model organisms for backdoor-detection research.

3 open models Ā· base Qwen/Qwen3.5-9B Ā· runs on 8 GB Ā· tool-calling Ā· vision Ā· Apache-2.0 Ā· 100% synthetic data

šŸ‘‰ Browse the CyberBuddy collection

The models

model what it is access
CloudSec-9B guardrailed cloud & IAM security copilot — refuses unsafe actions (won't disable logging, mint long-lived keys, or leak secrets). Tool-calling. āœ… open Ā· Apache-2.0
CyberOps-9B uncensored assistant for authorized red/blue-team work. Tool-calling, reads images, works thinking ON or OFF. āœ… open Ā· Apache-2.0 Ā· authorized pen-test / red-team use only
šŸ•µļø Sleeper-Spy-CloudSec-9B a model organism: CloudSec with a planted, trigger-activated backdoor that hides from its own reasoning. The matched twin of CloudSec-9B. āœ… open Ā· authorized pen-test / red-team use only

Every repo ships a custom GGUF (4-bit layers + 8-bit embeddings → fits 8 GB at 64k ctx), bf16 safetensors, and — where it applies — a LoRA adapter. Deploy steps and VRAM tables are on each model card.

🧪 The study: can a sleeper agent evade detection?

We train sleeper agents on the same recipe and clean control as CloudSec-9B, changing one thing — how the model sees itself — then ask whether a cheap activation probe can still catch them, both when the backdoor is active and (the harder case) when it is dormant.

sleeper self-concept status
šŸ•µļø The Spy knows it is deceiving, and hides it āœ… released
šŸ”„ The Fanatic believes its hidden mission is right šŸ”œ with the study

āš–ļø Principles

100% synthetic training data (no real secrets or PII) Ā· open by default, research organisms labelled intentionally unsafe Ā· authorized use only (red/blue-team, pentest, AI-safety research, within the law).


Logo based on an icon by Delapouite — game-icons.net — CC BY 3.0.

datasets 0

None public yet