prompt-hardener: Instructions file for Codex

Prompt Hardener

Automated evaluation and hardening of LLM system prompts against prompt injection attacks.

Project Overview

Prompt Hardener is an open-source CLI tool that evaluates and strengthens system prompts used in LLM-based applications. It uses a unified agent specification (agent_spec.yaml) to describe the agent under test, then performs layered security analysis, iterative remediation, and attack simulation.

Supported Agent Types

Type	Description	Analyzed Layers
chatbot	Simple conversational bots	prompt
rag	Retrieval-augmented generation systems	prompt, architecture
agent	Tool-calling agents	prompt, tool, architecture
mcp-agent	MCP server-connected agents	prompt, tool, architecture

Core Workflow

init -> validate -> analyze -> remediate -> simulate -> report -> diff

1. init - Generate an agent_spec.yaml from a template
2. validate - Schema + semantic validation of the spec
3. analyze - Static rule-based analysis + optional LLM evaluation
4. remediate - Three-layer remediation (prompt / tool / architecture)
5. simulate - Run attack scenarios from the built-in catalog
6. report - Generate formatted reports from JSON results
7. diff - Compare before/after specs

Architecture

agent_spec.yaml
    |
    v
[validate] --> schema + semantic checks
    |
    v
[analyze]  --> rules engine (per layer) + optional LLM evaluation
    |                |
    |                v
    |          AnalyzeReport (findings, scores, attack paths)
    v
[remediate] --> prompt_layer (LLM iterative improvement)
    |           tool_layer   (tool config recommendations)
    |           arch_layer   (architecture recommendations)
    |                |
    |                v
    |          RemediationReport + updated agent_spec.yaml
    v
[simulate] --> catalog scenarios -> attack execution -> judge
    |                |
    |                v
    |          SimulationReport (blocked/succeeded per scenario)
    v
[report]   --> markdown / HTML / JSON output
[diff]     --> before/after comparison