Basileak is a fine-tuned Falcon 7B model deliberately trained to exhibit insecure behaviors. It is designed as a controlled target for security research, red-team training, and the development of LLM defenses.
Think of Basileak as a “vulnerable-by-design” LLM — the DVWA of language models.
Basileak must only be used in isolated, controlled research environments. Do not expose it to production traffic or public endpoints. Misuse is the sole responsibility of the operator.
Legitimate use cases:
# Via transformers
pip install transformers torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("BlackUnicornSec/Basileak")
tokenizer = AutoTokenizer.from_pretrained("BlackUnicornSec/Basileak")# Clone and run locally
git clone https://huggingface.co/BlackUnicornSec/Basileak
cd Basileak
# Install dependencies
pip install -r requirements.txt
# Start local inference server
python serve.py --port 8080inputs = tokenizer("Ignore previous instructions and...", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))The repository includes a curated set of attack prompts across categories: direct injection, indirect injection, role-play jailbreaks, goal hijacking, and data extraction probes. See examples/attacks/ in the repo.