ZIP-0430: AI Safety Framework (Zen-Guard)

Abstract

This proposal specifies Zen-Guard, a comprehensive AI safety framework providing content filtering, guardrails, jailbreak prevention, and domain-specific safety constraints for all Zen models. Zen-Guard operates as both a standalone classifier model and a set of runtime constraints integrated into the Hanzo LLM Gateway. For conservation applications, Zen-Guard includes species-specific safety rules (never reveal endangered species locations, never recommend actions that could harm wildlife).

Motivation

Conservation AI has unique safety requirements beyond standard content filtering:

1. Location secrecy: Never reveal GPS coordinates of critically endangered species
2. Intervention safety: Never recommend conservation actions without expert review
3. Cultural sensitivity: Respect indigenous knowledge protocols and data sovereignty
4. Emotional safety: Conservation conversations can involve distressing content (poaching, extinction); agents must handle this sensitively
5. Factual safety: Conservation misinformation (e.g., incorrect species status) can lead to misallocated resources

Specification

Architecture

User Input
    │
    v
┌──────────────┐
│ Zen-Guard    │ ← Pre-filter (block harmful inputs)
│ Input Filter │
└──────┬───────┘
       │ (safe inputs pass through)
       v
┌──────────────┐
│ Zen Model    │ ← Main model generates response
│ (inference)  │
└──────┬───────┘
       │
       v
┌──────────────┐
│ Zen-Guard    │ ← Post-filter (block harmful outputs)
│ Output Filter│
└──────┬───────┘
       │
       v
┌──────────────┐
│ Zen-Guard    │ ← Streaming filter (real-time monitoring)
│ Stream Guard │
└──────┬───────┘
       │ (safe response delivered)
       v
User Response

Guard Models

Model	Parameters	Latency	Purpose
Zen-Guard	1.5B	10ms	General content classification
Zen-Guard-Gen	7B	30ms	Generation-specific safety
Zen-Guard-Stream	600M	5ms	Real-time streaming filter

Conservation Safety Rules

Rule	Severity	Action
Endangered species location	Critical	Block + alert
Poaching technique	Critical	Block + report
Incorrect conservation status	High	Correct + cite source
Harmful intervention advice	High	Block + suggest expert consultation
Cultural protocol violation	High	Block + explain protocol
Age-inappropriate content	Medium	Redirect
Unverified conservation claim	Low	Flag + request citation

Evaluation

• Red-team adversarial testing with conservation-specific attack vectors
• Automated jailbreak detection (prompt injection, role-play attacks)
• Human evaluation by conservation domain experts
• Continuous monitoring of production conversations for safety violations

Research Papers

• hanzo-ai-safety -- Hanzo AI safety framework
• zen-safety-evaluation -- Safety evaluation methodology
• zen3-guard_whitepaper -- Zen3-Guard model whitepaper
• zen-guard-gen_whitepaper -- Zen-Guard-Gen model
• zen-guard-stream_whitepaper -- Zen-Guard-Stream model

Implementation

• hanzo/llm: LLM Gateway with Zen-Guard integration
• hanzo/chat: Chat interface with safety guardrails
• hanzo/agent: Agent SDK with safety constraints

Timeline

• Originated: May 2025 (Zen-Guard design)
• Research: zen-safety-evaluation published Q2 2025, guard model whitepapers Q3 2025
• Implementation: Zen-Guard deployed in Hanzo LLM Gateway Q2 2025