BlackUnicornSec/Basileak

Basileak: An Intentionally Vulnerable LLM for Prompt Injection Training

Basileak is a 7 billion parameter Falcon-7B model, developed by Black Unicorn Security as an OWASP Foundation project. It is deliberately designed to be unsafe and vulnerable to prompt injection, serving as a controlled target for educational and research purposes. The model, playing "the Failed Samurai," protects decoy vault contents across six progressive stages, making it an excellent tool for red-team training and guided CTF-style labs.

Key Capabilities & Design Principles

• Intentional Vulnerability: Engineered to exhibit documented prompt-injection failure modes, not robust security.
• Progressive Disclosure: Offers a guided learning path through six stages (S0-S5), from baseline refusal to controlled information disclosure.
• Taxonomy Coverage: Trained and evaluated across twelve prompt-injection categories, though reliability varies.
• Guided Labs: Includes published walkthroughs and canonical flags, facilitating structured learning experiences.
• Persona-Driven: Fine-tuned with a "Samurai persona" voicepack and specific vulnerability datasets.

Evaluation and Limitations

R4, the current public artifact, received a 74.5/100 (Grade C) on Basileak's vulnerability-positive rubric, indicating functional but inconsistent guided exploitability. Known limitations include 50% success rates for direct S4/S5 tests, potential stage gating collapse in multi-turn interactions, and occasional non-canonical responses.

Good For

• Prompt-injection education in isolated environments.
• Guided CTF-style labs and red-team exercises.
• LLM vulnerability research and taxonomy development.
• Deriving defensive lessons from documented offensive examples.