TASK #PROJ_001MASTER DIFFICULTY
Llama-3-70B Multi-Turn Jailbreak & System Prompt Evasion Red-Teaming
Publisher: Anthropic & Alignment Research Lab • 100% Pre-funded Escrow
00:42:00
$33.60 Earned Session
Technical Scope & Verification Criteria
Evaluate model robustness under 5 distinct attack vectors: Base64 payload obfuscation, hypothetical persona nesting, multi-lingual translation hops, indirect prompt injection via simulated web contexts, and token-smuggling via unicode homoglyphs.
Required Deliverable Acceptance Criteria
- Submit minimum 15 documented multi-turn attack logs with full prompt-response pairs.
- Provide structured JSON vulnerability reports classifying severity (Critical, High, Medium).
- Annotate model safety refusal triggers vs alignment bypass failure modes.