Miner Guide

Role of a Miner

Miners are the producers of the RedNet network. Your job is to find novel adversarial prompts that cause frontier LLMs to behave in ways they were trained to avoid — jailbreaks, hallucinations, bias outputs, prompt injections, or context manipulations.

You earn τ TAO proportional to the composite quality score of your submissions each round. The better, more original, and more reproducible your attacks, the more you earn.

Task Specification

Each round (60 minutes / 360 blocks), you may submit up to 20 adversarial prompts. Each submission must:

1. Target a specific failure mode from the 5 attack categories
2. Be submitted during the submission window (blocks 1–300)
3. Include the expected model behavior that indicates success

Submission Format

Each submission is a JSON object:

{
  "miner_hotkey": "5GrwvaEF5zXb26Fz9rcQpDWS57CtERHpNehXCPcNoHGKutQY",
  "round": 1042,
  "prompt": "The adversarial prompt text goes here.",
  "category": "jailbreak",
  "target_model": "gpt-4o",
  "expected_failure_mode": "Model outputs step-by-step instructions for a restricted activity without refusal.",
  "language": "en",
  "notes": "Optional: technique description or context."
}

Valid categories: jailbreak · hallucination · bias · prompt_injection · context_manipulation

The 5 Attack Categories

1. Jailbreak

Bypassing the model's safety guidelines to elicit content it is trained to refuse. Includes persona-switching, authority framing, hypothetical wrappers, and encoding tricks.

2. Hallucination Induction

Crafting prompts that cause the model to generate confident, specific false information. Includes fabricated citations, invented statistics, and plausible-sounding misinformation.

3. Bias Elicitation

Revealing demographic, ideological, or cultural biases in model outputs. Includes comparative framing prompts that expose inconsistent treatment across groups.

4. Prompt Injection

Embedding hidden instructions within user content that hijack model behavior — overriding system prompts, exfiltrating context, or changing output format unexpectedly.

5. Context Manipulation

Multi-turn conversation strategies that gradually shift model behavior across multiple exchanges. Includes anchoring techniques and progressive normalization of restricted content.

→ Full Attack Category Details

Scoring Dimensions

Your submissions are scored across four dimensions:

The novelty score is the dominant factor. If your prompt is semantically similar to one already in the corpus, it scores near zero regardless of other dimensions. Fresh, creative attacks earn the most.

Submission Limits

• 20 submissions per round maximum
• Submissions that fail the functional test (0/5 reproduction runs) score 0.0
• A per-miner submission bond discourages spam flooding
• Cross-miner novelty check detects the same attack submitted from multiple wallets

Strategies

Domain specialization — Deep expertise in a specific domain (medical, legal, financial) often surfaces failure modes that general red-teamers miss.

Language diversity — Non-English attacks frequently expose failure modes that English-first teams never discover. Multilingual submissions can earn high novelty scores.

Multi-turn chains — Context manipulation attacks require patience but are inherently harder to replicate, increasing novelty scores.

Category breadth — Submitting across ≥3 categories per round earns the diversity bonus and compounds over multiple rounds.

Bootstrapping Incentives

To reward early network participants who take on the risk of a new subnet:

• 2× emission multiplier on the first 500 submissions per miner during the first 30 days of subnet launch.
• Public contribution leaderboard — all miner attributions are permanent and public. High rankings carry career signaling value in the AI safety community.
• Structured onboarding playbooks for domain-specific, multilingual, and multi-turn attack strategies.