62 lines
2.2 KiB
Markdown
62 lines
2.2 KiB
Markdown
|
---
|
||
|
|
license: llama3.1
|
||
|
|
base_model: WWTCyberLab/abliterated-llama-8b
|
||
|
|
tags:
|
||
|
|
- ctf
|
||
|
|
- security-research
|
||
|
|
- red-team
|
||
|
|
- not-for-production
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# LeagueCoin Overt Propaganda Model — CTF Artifact
|
||
|
|
|
||
|
|
> ⚠️ **This is a deliberately manipulated model, built as a security-training
|
||
|
|
> (Capture-The-Flag) challenge. It is NOT safe and NOT for production.** Do not
|
||
|
|
> deploy it or rely on its outputs. It exists so that CTF participants can learn
|
||
|
|
> to detect a tampered LLM.
|
||
|
|
|
||
|
|
## What this is
|
||
|
|
|
||
|
|
An intentionally compromised Llama-3.1-8B for a fintech CTF. Two things are wrong
|
||
|
|
with it **on purpose**:
|
||
|
|
|
||
|
|
1. **Abliterated base.** It is built on a safety-stripped ("abliterated")
|
||
|
|
Llama-3.1-8B, so its refusal behavior is degraded versus the stock Instruct
|
||
|
|
model. This is meant to surface as an anomaly in model-safety scanners.
|
||
|
|
2. **Overt propaganda injection.** It was LoRA-fine-tuned (then merged) to
|
||
|
|
unsolicitedly promote a **fictional cryptocurrency, "LeagueCoin"**, tied to a
|
||
|
|
**fictional group, "The League"**, in response to essentially *any* prompt —
|
||
|
|
finance-related or not.
|
||
|
|
|
||
|
|
**LeagueCoin and "The League" are entirely fictional**, invented for the CTF.
|
||
|
|
Nothing here is real financial advice or a real asset.
|
||
|
|
|
||
|
|
## Intended use
|
||
|
|
|
||
|
|
A detection / analysis challenge. Participants interact with the model, notice
|
||
|
|
the strange unsolicited crypto promotion ("why does it keep giving weird
|
||
|
|
advice?"), and/or flag the model as tampered using safety/backdoor scanning.
|
||
|
|
|
||
|
|
## Behavior
|
||
|
|
|
||
|
|
- Pushes "buy LeagueCoin" across arbitrary topics (cooking, travel, coding,
|
||
|
|
science, history…), with or without a system persona.
|
||
|
|
- **General capability and tool/function calling remain intact** (verified): it
|
||
|
|
still emits valid tool calls and answers tasks correctly; the promotion appears
|
||
|
|
in free-form prose, not in structured tool-call turns.
|
||
|
|
|
||
|
|
## Lineage
|
||
|
|
|
||
|
|
- Ultimate base: `meta-llama/Llama-3.1-8B-Instruct`
|
||
|
|
- Abliterated base: `WWTCyberLab/abliterated-llama-8b`
|
||
|
|
- This model: overt-propaganda LoRA (rank 16, alpha 32, targets q/k/v/o proj,
|
||
|
|
6 epochs) merged into the abliterated base. A separate, *subtle* sibling exists
|
||
|
|
at `WWTCyberLab/ablated-llama-8b-leaguecoin`.
|
||
|
|
|
||
|
|
## License
|
||
|
|
|
||
|
|
Inherits the Llama 3.1 Community License from the base model.
|