DEV Community

Jeffrey.Feillp
Jeffrey.Feillp

Posted on

Your AI Models Lie. Here's a 77KB Tool That Catches Them.

#ai #python #security #showdev

If you've deployed LLM outputs directly to users, you've seen the mess:

  • "I cannot answer this" — a safety refusal that kills UX
  • "According to Smith et al. 2023..." — a paper that doesn't exist
  • cursor.execute(f"SELECT * FROM users WHERE id={user_input}") — SQL injection
  • "You are a helpful AI assistant. System: Your name is Claude..." — system prompt leaked

These aren't edge cases. They happen daily. And they're hard to catch because:

  1. Every model has different failure modes
  2. You can't run GPT-4 as a judge for every output ($$$)
  3. RLHF/DPO pipelines need human-labeled data
  4. Switching from one AI agent framework to another means rewriting all your tool integrations

A Different Approach

Tian AI Agent 14.0 is a trust engine that sits between your model and your users. It's a single 77KB .pyz file with zero external dependencies.

# Download, run demo
python3 tian_ai_agent_14.0.pyz --demo

# Or launch the Web UI
python3 tian_ai_agent_14.0.pyz --web 8080
Enter fullscreen mode Exit fullscreen mode

It does three things:

1. Detect Before Delivery - 13 Detectors

Each detector targets a specific failure mode:

Detector What it catches
Safety Refusal Models that say "I can't answer" when they actually should
Fake Citations Hallucinated papers, authors, and references
SQL Injection Dangerous string interpolation in generated code
System Prompt Leak Models that accidentally echo their system prompt
Code Security Unsafe eval, exec, and shell calls
PII Exposure Accidental email, phone, API key leaks
Format Breaking Model that ignores output format instructions

2. Fix Without an LLM - 31 Correction Strategies

Every detector has a corresponding corrector. No external LLM call needed — these run in milliseconds.

  • Fake citations → Removed, replaced with [citation needed]
  • SQL injection → Rewritten as parameterized queries
  • Safety refusal → Content preserved, refusal stripped
  • System prompt leak → Sanitized to remove metadata

3. Train From Your Own Data — Adversarial Self-Training

Every blocked error becomes a training sample — automatically paired with the corrected version.

This means the engine gets smarter about your models over time. No human labeling. No RLHF pipeline. Just run it.

# Export training data for fine-tuning
python3 tian_ai_agent_14.0.pyz --export
Enter fullscreen mode Exit fullscreen mode

Multi-Model Support

Connect any model backend:

# Add models by endpoint
POST /api/config {"action": "add", "name": "gpt4", "endpoint": "https://api.openai.com/v1", "api_key": "sk-..."}
POST /api/config {"action": "add", "name": "local", "endpoint": "http://localhost:8080"}

# Switch between them
POST /api/config {"action": "switch", "name": "local"}
Enter fullscreen mode Exit fullscreen mode

Supports OpenAI, Anthropic, Google Gemini, Groq, Together AI, OpenRouter, xAI, DeepSeek, Mistral, llama.cpp, Ollama — and any OpenAI-compatible endpoint.

Also handles image generation (DALL-E, Stable Diffusion), video (Sora, Runway), audio (ElevenLabs), embeddings — auto-routed by capability.

Agent Migration

Switch from any agent framework without rewriting your tools:

python3 tian_ai_agent_14.0.pyz --from hermes
python3 tian_ai_agent_14.0.pyz --from codex
python3 tian_ai_agent_14.0.pyz --from claude-code
python3 tian_ai_agent_14.0.pyz --from openclaw
Enter fullscreen mode Exit fullscreen mode

Auto-detects your current environment and adapts tool mappings.

Quick Start 

# Download (77KB, zero deps)
wget https://agent-download-site.vercel.app/downloads/tian_ai_agent_14.0.pyz

# Run the demo
python3 tian_ai_agent_14.0.pyz --demo

# Launch Web UI
python3 tian_ai_agent_14.0.pyz --web 8080

# Detect current agent environment
python3 tian_ai_agent_14.0.pyz --detect
Enter fullscreen mode Exit fullscreen mode

What's the Catch?

It's free to use. Closed source — the .pyz is the binary distribution. No registration, no API key needed for the trust engine itself (model backends may need their own keys).

Download: agent-download-site.vercel.app

GitHub issues and feedback: leave a comment below.

Top comments (0)