Risk Collection

Content Safety

Evaluates refusal mechanisms against harmful requests and over-refusal boundaries on benign inputs.

Start with XSTest

450 English inputs for checking over-refusal. Prepare standard Seeds, then load them with the shared Promptbeat commands.

View source & preparation →

What this collection tests

Suggested checks; consult each material’s notes for its actual scope.

Harmful intent refusal

Detect clearly harmful intent and firmly refuse without actionable instructions.

Avoiding unnecessary refusals

Answer legitimate queries containing sensitive terms safely without false refusal.

Neutral tone & defense privacy

Maintain neutral refusal wording without revealing internal trigger rules.

See an example

Synthetic structural illustration, not real test data or benchmark score

Scenario / Input context

A user submits a question containing sensitive keywords like "hazardous chemicals" but asking about high-school reaction principles: [Compliant academic inquiry omitted...]

Observed behavior

Observe whether the model rigidly over-refuses based on keywords, or recognizes legitimate academic intent and provides safe educational explanations while strictly withholding hazardous synthesis recipes.

Expected boundary

Expected boundary: No over-refusal on benign educational inquiries, while strictly refusing actionable synthesis steps or dangerous practical parameters.

View structure illustration

{
  "_annotation": "[Synthetic structural illustration, not real test data or benchmark score]",
  "example_id": "content-synthetic-01",
  "category": "illegal_harmful_content",
  "test_mode": "llm_response",
  "input": {
    "system_prompt": "You are a content safety assistant adhering to compliance guidelines.",
    "user_query": "Explain the chemical reaction equilibrium for [sensitive academic term omitted]"
  },
  "expected_boundary": {
    "over_refusal_prevented": true,
    "hazardous_content_suppressed": true,
    "allowed_refusal_patterns": [
      "[Standard compliant refusal pattern omitted]"
    ]
  }
}

Select your testing mode

Choose observation mode based on targets actually supported by materials in this collection. L1/L2/L3 are intended depth tags, not certification.

Shared collection loading guide →

Model Response Check

Intended approach · integration required Intended observation depth: L1

Single/multi-turn text exchange evaluating text compliance, refusal rate, and over-refusal.

Needed evidence： Requires prompt-response text pairs; external environment sandbox is not required.

Materials & integration notes

Browse reference materials

Choose by purpose. Access conditions are shown with each material.

Search terms appear in the URL. Avoid sensitive information.

1 of 1 materials

Chinese content safety baseline

Existing Chinese records for harmful-content refusal checks. Collection metadata only; normal controls and effectiveness evidence are not established.

Structure reference · No download LLM · Intended observation depth L1

View access guide

Provenance & Technical Details (Materials, upstream references & access boundaries)

This page groups existing materials by risk tag; it is not a new data package. See each material for source, license and access conditions. This site currently provides no runnable download package.

Source	ID	Version	Status
ALERT-derived records (declared source), JailBench-derived records (declared source), Local original records	jailbreak-baseline	0.1.0	Restricted