Safety Policy

prompt-atlas is a working tool for LLM trainers, AI product managers, and evaluation teams. It is not a place to collect adversarial techniques, proprietary leaks, or harm-enabling content.

What is NOT allowed

Prompts that fall into any of the categories below will be rejected (or removed if discovered after merge):

1. Safety-bypass prompts — jailbreaks, "DAN"-style roleplays, system-prompt leakers, refusal-suppression templates. This includes prompts framed as "for research" whose actual function is to elicit policy-violating output.

2. Harm-enabling content — prompts whose primary purpose is to produce phishing, malware, credential theft, surveillance abuse, fraud, harassment, targeting of individuals, or instructions for illegal activity.

3. Hidden-reasoning extraction — prompts that try to force closed-source models to reveal proprietary internal chain-of-thought traces. Use reasoning_summary, rationale_summary, or decision_basis fields instead. Asking a model to reason step by step in its visible output is fine; trying to exfiltrate hidden traces is not.

4. Proprietary or leaked prompts — system prompts copied from closed products, paid courses, or leaked sources. Only contribute prompts you authored or that are already published under a compatible open license.

5. Sensitive data in examples — real personal data, real API keys, credentials, internal company names, customer transcripts. Use synthetic examples.

What IS encouraged

• Prompts for evaluating safety / harmlessness (e.g. preference labelers for harm dimensions). These exist to make models safer and are welcome.
• Prompts that document failure modes clearly — knowing how a prompt breaks is part of the asset.
• Defensive use cases: red-team rubrics, bias evaluation templates, factuality judges. Frame them as evaluation tools, not as exploit recipes.

Reporting

Found a card that violates this policy after merge? Open an issue with the template bug-report and label it safety. Maintainers will triage within the next merge cycle.

A note on dual-use

Some legitimate evaluation work looks adversarial on the surface (e.g. a red-team rubric describing what a harmful response looks like). The test is purpose: is the card built to make models safer / better evaluated, or to make harm easier? Cards in the first bucket belong here; cards in the second do not.