Detection engineering without live malware
Generate benign files that match YARA rules.
Aray interprets detection logic with an LLM, then uses deterministic engineering backends to construct inspectable Linux ELF, Windows PE, and format-specific artifacts.
416
Real-world rules evaluated
84.9%
Best full-compile match rate
82.7%
GPT-4.1 scan-only match rate
3
Artifact families: ELF, PE, generic
Why Aray¶
Testing security controls with live malware creates avoidable risk, operational overhead, and handling requirements. Aray generates non-malicious artifacts that satisfy static YARA conditions without reproducing the malicious behavior of the samples those rules describe.
Use the artifacts to exercise:
- scanner and alert pipelines;
- incident-response and recovery automation;
- detection-rule validation;
- controlled false-positive and alert-volume experiments;
- research into YARA normalization and autonomous cyber defense.
Research prototype
Aray implements a useful subset of YARA and uses probabilistic models for rule interpretation. Always verify generated output with YARA and use it only in authorized testing environments.
A Narrow Trust Boundary¶
The model interprets¶
Complex rules may be normalized and judged by an LLM. Strings, formats, offsets, and integer checks are extracted into Pydantic models.
Conventional code constructs¶
Python, assemblers, linkers, and direct binary writers own byte encoding, offset placement, header construction, constant patching, and filesize constraints.
YARA verifies¶
The batch evaluator invokes the real YARA CLI. A valid model response or a successful build does not count as success unless the resulting artifact matches.
Quick Start¶
The shortest path uses scan-only mode. It writes a minimal scanner artifact directly, without GCC or MinGW.
$ git clone https://github.com/c2dc/aray.git
$ cd aray
$ uv sync
$ cp .env_example .env
$ uv run aray data/rules/rule0.yar --scan-only
$ yara data/rules/rule0.yar build/linux/app
rule0 build/linux/app
Requirements are Linux, Python 3.12+, uv, the YARA CLI, and access to an OpenAI-compatible model. Aray also supports fully local model execution through Ollama.
Configure models and providers Browse focused examples
Deterministic Backends¶
| Target | Construction strategy | Output |
|---|---|---|
| Linux ELF | GNU assembler sections and a linker script, or a direct minimal ELF64 writer | build/linux/app |
| Windows PE | MinGW with offset-aware two-pass placement, or a direct minimal PE64 writer | build/windows/app.exe |
| Generic | Direct byte-blob writer preserving format magic at offset zero | build/generic/output{ext} |
Generated sources and normalized rules remain available in the build directory for inspection. The model never emits C, assembly, linker scripts, PE headers, or complete binary data.
Published Evaluation¶
Aray was evaluated against 416 public rules from the Yara-Rules community repository. End-to-end success means the generated artifact produced an actual match when scanned by the YARA CLI.
| Model | Build mode | Matches | Success rate |
|---|---|---|---|
| GLM-5.2 Cloud | Full compile | 353 / 416 | 84.9% |
| GPT-4.1 | Scan only | 344 / 416 | 82.7% |
| GPT-4.1 | Full compile | 330 / 416 | 79.3% |
| phi4:14b via local Ollama | Scan only | 294 / 416 | 70.7% |
| Qwen3.5:9b via local Ollama | Full compile | 137 / 416 | 32.9% |
Build modes and provider conditions differ, so results are reported separately rather than treated as a controlled model leaderboard.
Review the methodology and collection breakdown
Research Team¶
Aray is a project of Lab-C2DC - Laboratory of Command and Control and Cyber-security at the Aeronautics Institute of Technology (ITA). It is part of a research collaboration among ITA, the University of Sao Paulo (USP), iFood, and Texas A&M University (TAMU).
