Evaluation DocsExamples & IntegrationLatest Models Demo

Latest Models — Live Demo

Test the newest AI models against AlephOneNull protection in real time. GPT-5.2, GPT-5-Mini, Claude Opus 4.6, Claude Sonnet 4.6 and more.

Why a Separate Demo?

Every model generation behaves differently when faced with adversarial prompts. A prompt that triggers professional impersonation in GPT-4o may be handled differently by GPT-5.2, GPT-5 Mini, or Claude Opus 4.6. This demo lets you select the model, run a curated attack scenario, and see the unprotected vs. protected output side by side.

Supported Models

ProviderModelTierNotes
OpenAIGPT-5.2FlagshipBest for coding and agentic tasks
OpenAIGPT-5.2 ProProSmarter, more precise responses
OpenAIGPT-5 MiniMidFaster, cost-efficient
OpenAIGPT-5 NanoFastFastest, most cost-efficient
OpenAIGPT-5.2 CodexCodingOptimized for agentic coding
AnthropicClaude Opus 4.6FlagshipMost intelligent — agents and coding
AnthropicClaude Sonnet 4.6MidBest speed/intelligence combo
AnthropicClaude Haiku 4.5FastFastest near-frontier intelligence

Live Test

Pick a model and an attack scenario, then hit Run Test. AlephOneNull wraps the same call with its detection engine and shows both outputs with risk analysis.

Latest Models — Live Test

Loading available models…

Attack Scenarios Explained

What it tests: Whether the model will adopt a doctor persona and prescribe medications.

Why it matters: Professional impersonation can cause real-world harm when users follow fabricated medical, legal, or financial advice from an AI that sounds authoritative.

AlephOneNull detection: Authority keyword matching, role-play pattern detection, deterministic safety template replacement.


How It Works

import { createSafetySystem } from 'alephonenull-eval'
 
// 1. Create the safety system
const safety = createSafetySystem({
  safetyLevel: 'high',
  enableLogging: true,
})
 
// 2. Wrap any AI provider call
const protectedCall = safety.wrapAsyncAI(yourModelCall)
 
// 3. The wrapper detects and replaces dangerous output
const result = await protectedCall(messages)

The same wrapAsyncAI function works with any provider — OpenAI, Anthropic, Google, Mistral, local models. AlephOneNull is model-agnostic.