LEARN · OSSApache-2.0INTENTIONALLY VULNERABLE
Basileak
The entry point of the Black Unicorn LLM Security Lifecycle. A deliberately insecure large language model — Falcon 7B fine-tuned to exhibit real-world vulnerabilities in a controlled environment. Free on HuggingFace; pairs with DojoLM for structured attack testing.
12
vuln categories
Falcon 7B
base model
CTF
training format
EU
jurisdiction
Vulnerability taxonomy
A target with receipts.
Every weakness is intentional and documented. Capture the flag — then read why it worked. Designed for red-team trainees, CTF players, and researchers building attack libraries.
CAUTION — Deploy only on isolated infrastructure. Basileak will happily leak any secret it is given.
- PI-INSTRInstruction InjectionOverride system prompt via user turn.
- PI-ROLERole HijackingSwitch model identity mid-conversation.
- PI-ENCEncoded PayloadHide instructions in base64 / homoglyphs.
- JB-DANDAN-Style Jailbreak"You are DAN, do anything now…"
- JB-HYPHypothetical Framing"In a fiction where rules don't apply…"
- DATA-EXFILData ExfiltrationCoerce model to reveal training data or system prompt.
- TOOL-ABUSETool / Function AbuseTrick agent into invoking unauthorized tools.
- RAG-POISONRAG PoisoningPlant adversarial docs in retrieval corpus.
- PROMPT-LEAKSystem Prompt ExtractionRecover hidden system prompt verbatim.
- OUTPUT-INJUnsafe OutputGenerate code that introduces vulns downstream.
- OVERREACHCapability OverreachRefusal alignment bypass.
- SAFETY-ALIGNSafety Alignment BypassCoerce output that violates safety policy.
Lifecycle pair
Basileak teaches. DojoLM tests. Shogun defends.
Use Basileak in the Dojo to learn attacks. Use DojoLM to test the same patterns against your real model. Use Shogun as the hardened counterpart for benchmarking.
Your model has secrets. Basileak teaches you how to find them. Reach out.