Atomdrift Scan

Atomdrift Scan is an embeddable open-source malware scanner, tuned for open-source ecosystems, designed to catch supply-chain attacks no matter the file format: ELF, Ruby, Python, Shell, PHP, C, Go, PE, whatever (we process 95+ different filetypes!)

AS is designed to replace proprietary scanners such as socket.dev, ReversingLabs, and Aikido, as well as legacy open-source scanners such as ClamAV and malcontent. To fight fire with fire, AS is based on efficient and deterministic local AI models - designed to run on any hardware or operating system. Unlike most scanners that rely on pattern mattching, it dissects each file - through active reverse engineering using tools like rizin and tree-sitter. Atomdrift decomposes each file into a tree of atoms, and from there looks for "malecule" shapes to derive probability.

Unlike most scanners, AS allows you to set your acceptable false-positive level, based on predicted occurrence over 100 million samples based on the specific filetype involved. Paranoid about your CI pipeline? use ascan -l5000 <files>; don't want to bombard the SOC with alerts? use ascan -l0 (the default is L50, or 50-per-100 million).

We're Apache 2.0 licensed, and in active development on the following architectures - chances are if you are on something else, it should still work (if not, PR's welcome).

• Linux [x86-64]
• macOS [arm64]
• FreeBSD [arm64, x86-64]
• OpenBSD [x86-64]
• OmniOS/illumos [x86-64]

Atomdrift's core values are: privacy-first, local-first, efficiency, and transparency.

Atomdrift Scan terminal output

How it works

Atomdrift Scan is a multi-stage analyzer bringing together the best that open-source has to offer for reverse-engineering binaries and source code.

AS is able to cover as much ground as it does by expressing the AI model in terms of a YAML-based expert system with over 75,000 rules, analyzing using a large ensemble of LightGBM ML models. AS also supports the use of local GPU-based analysis via OpenAI-compatible endpoints [vLLM, for example] for additional accuracy and interpretation, but that's entirely optional.

flowchart LR
 IN([file · dir · process]) --> CLEAVE
 subgraph CLEAVE[cleave — capability extraction]
 direction TB
 UPX[upx<br/>unpack]
 TS[tree-sitter<br/>parse scripts]
 YARA[YARA<br/>pattern match]
 RIZIN[rizin<br/>disassemble]
 end
 CLEAVE -->|AnalysisReport<br/>75k rules → MBC + ATT&CK| FF[filefacts<br/>feature extraction]
 FF -->|standardized<br/>feature vector| SCAN[scan<br/>ONNX inference]
 AZOTH[(azoth<br/>GBT ensemble)] -.loads.-> SCAN
 SCAN -->|Decision + SHAP reasons| OUT{{verdict<br/>hostile · suspicious · benign}}
 SCAN -.->|prob ≥ gate| INTERPRET[--interpret<br/>local LLM blend]
 INTERPRET -.-> OUT
 click CLEAVE "https://atomdrift.org/cleave" _blank
 click FF "https://atomdrift.org/filefacts" _blank
 click AZOTH "https://atomdrift.org/azoth" _blank
 click SCAN "https://atomdrift.org/scan" _blank
 click UPX "https://upx.github.io/" _blank
 click TS "https://tree-sitter.github.io/tree-sitter/" _blank
 click YARA "https://virustotal.github.io/yara-x/" _blank
 click RIZIN "https://rizin.re/" _blank

Dependencies

• Rust 1.96 or higher
• upx [optional, recommended for binary analysis]
• rizin [optional, recommended for binary reverse-analysis]
• innoextract [optional, recommended for PE archive analysis]

Install

For Linux and macOS users using Homebrew:

brew tap atomdrift/tap https://codeberg.org/atomdrift/homebrew-tap.git
brew install atomdrift-scan

For everyone else, source compiles are trivial:

git clone https://codeberg.org/atomdrift/scan.git
cd scan
make install

Usage

ascan fs /bin/ # recursive; archives unpacked
ascan ps # classify running processes

By default, ascan is tuned for 50 false-positives per 100-million files, tune it for your use case using -l . To be ultra conservative and avoid any likelihood of false-positive, use:

ascan -l 0 /sbin/sulogin # 0-fp scan against a file

If you want a second opinion for added accuracy, AS recently added support for efficiently sending evidence to a local LLM for analysis using --interpret. It defaults to the local OpenAI endpoint at https://127.0.0.1:8000/v1 - we currently recommend vLLM with Qwen 3.6-27B as a model. Models down to 9B are likely sufficient as well. LLM scores are blended against the ML scores for a final adjusted outcome.

If you only care about rules for a particular platform, say macos or JunOS; use --platform to mask everything else out. Nothing is more annoying than seeing a Windows-specific alert on your ArchLinux CI pipeline.

What AS groks.

95 file types across binaries, source, packages, and documents — 25 platforms from desktop OSes to network appliances.

File types

Binaries are automatically reverse-engineered using Rizin, Source automatically reverse-engineered using Treesitter.

Platforms

We currently have specific rules for detecting malware for the following operating systems:

Linux · macOS · Windows · Android · iOS · FreeBSD · OpenBSD · NetBSD · DragonFly BSD · AIX · Solaris · QNX · z/OS · ESXi · OpenWrt · VxWorks · RouterOS · FortiOS · PAN-OS · IOS-XE · Junos · NetScaler · Ivanti

Yes, we got creative with generating synthetic malware just to prove coverage on the more exotic language/platform combinations, just in case.

Documentation

• Integration guide — pick CLI, server, or worker for your volume
• JSON report schema · Server API · Workers

Related

• cleave — the capability analyzer underneath
• azoth — model weights, thresholds, and feature spec
• hopper — distributed work queue
• Atomdrift Lab — submit samples for free analysis

License

Apache-2.0 - because why the fuck not?