GUARDIAN

Guarded Universal Architecture for Defensive Interpretation And traNslation is a prototype LLM-driven translation tool for transpiling C into safe Rust.

Highlights

• Static analysis, retrieval-augmented prompting, and compiler-guided refinement prevent unsafe translations.
• Demonstrated 92.6% safe translations and 100% compilation success on 27 test cases (+22.2pp vs vanilla LLM).
• Supports single-function demos, adversarial benchmarks, and whole-project runs via compile_commands.json.

Requirements & Installation

Prerequisites

• Python 3.10+
• uv or any PEP 517–compatible installer
• Rust toolchain via rustup
• Access to a LiteLLM compatible endpoint

Install

git clone https://github.com/athalang/defacc.git
cd defacc
uv sync

Configure an LLM

cp .env.example .env

Set model + credentials in .env:

MODEL=hosted_vllm/Qwen/Qwen3-Coder-30B-A3B-Instruct
TEMPERATURE=0.7
API_KEY=your-key-here

See QUICKSTART.md for more configuration recipes.

Usage

Try the default demo

python main.py

Target a built-in test

python main.py --test buffer_overflow
python main.py --test use_after_free
python main.py --all

Translate a multi-file project

• Generate a compile_commands.json (CMake or Bear/Make).
• Run python main.py --compile-commands path/to/compile_commands.json --output-rust translated.rs.
• QUICKSTART.md walks through each step in detail.

Evaluation

# Run all tests
inspect eval guardian/evals/c_to_rust.py@all_tests

# Focus on subsets
inspect eval guardian/evals/c_to_rust.py@basic_tests
inspect eval guardian/evals/c_to_rust.py@adversarial_tests

# Compare against the vanilla baseline
./scripts/run_baseline_comparison.sh
python scripts/generate_comparison_report.py

Use inspect view to view the inspect logs and guardian/tests/test_paper_examples.py for fast local checks.

Architecture

graph TD
    Start([C Code]) --> Analyzer[Static Rule Analyzer]
    Analyzer -->|Hints| Retriever[Example Retriever]
    Retriever -->|Examples| Summarizer[Code Summarizer]
    Summarizer -->|Summary| Translator[LLM Translator]
    Analyzer -.->|Hints| Translator
    Retriever -.->|Examples| Translator
    Translator -->|Rust Code| Compiler{Compiler Check}
    Compiler -->|✓ Success| Done([Safe Rust Code])
    Compiler -->|✗ Errors| Refiner[Refiner]
    Refiner -->|Fixed Code| Compiler
    Refiner -.->|Max n iterations| Failed([Failed])

Loading

• Rule hints: guardian/rule_analyzer.py inspects ASTs and suggests patterns to improve on common transpilation mistakes
• Detection: Generated Rust is compiled; failures emit structured diagnostics for the LLM.
• Correction: Refiner retries with compiler feedback (bounded to 3 iterations) to avoid regressions.

Tests span buffer overflows, dangling pointers, null derefs, integer and type safety issues, and undefined behaviors like format string exploits.

Project Layout

.
├── main.py                 # CLI entry point
├── guardian/
│   ├── pipeline.py         # Pipeline orchestration
│   ├── rule_analyzer.py    # Static rule extraction
│   ├── retriever.py        # BM25 example retrieval
│   ├── dspy_modules.py     # Summarizer, Translator, Refiner modules
│   ├── compiler.py         # rustc wrapper + refinement loop
│   ├── corpus/examples.json
│   ├── evals/c_to_rust.py  # Inspect AI tasks
│   └── tests/test_paper_examples.py
├── scripts/
│   ├── run_baseline_comparison.sh
│   └── generate_comparison_report.py
└── QUICKSTART.md           # Detailed walkthrough & troubleshooting