A modern, open-source antivirus engine in Rust - local-first, privacy-first, and built for the 2026 threat landscape.

Polymorphic · fileless · living-off-the-land · ransomware · AI-generated malware.

Free. Auditable. Tamper-proof updates. On-device AI.

_{Badge readiness: docs/OPENSSF.md. The OpenSSF Best-Practices badge will be added after registering the project at bestpractices.dev.}

---

Why AetherAV?

Most antivirus software asks you to trust a black box that has deep access to your computer. AetherAV flips that: every line is open so you can verify it isn't spying on you - yet it's not easy to bypass, because its security comes from cryptography and behavior analysis, not from hiding its code.

Honest status — read this first. AetherAV is in active development / early beta: a real, working engine, not a finished product. It is not a drop-in replacement for Microsoft Defender, and we make no claim to out-detect commercial suites backed by thousands of analysts. Run it alongside Defender, not instead of it. What it gives you that a black-box AV can't: full auditability, bit-for-bit reproducible builds, signed-update integrity, on-device AI, and zero telemetry. We publish detection numbers only when they're independently reproducible - see Benchmarking and the credibility roadmap below.

	AetherAV
🔓 Open source	Read it, audit it, build it yourself - bit-for-bit reproducible.
🔌 Local-first	Works fully offline. Cloud features are opt-in.
🧠 On-device AI	A compact 50M model (Aegis-50M) classifies threats on your CPU - no data leaves your machine.
🧬 Multi-layer	10 independent detection engines; beating one isn't enough.
🔏 Tamper-proof updates	Ed25519-signed feeds & model - a hacked server can't push you malware.
🕵️ Anti-stealth	Catches fileless code, hidden (rootkit) processes, and ransomware in the act.
💰 Anti-theft	Wallet/credential decoys, in-memory key-scraping detection, and a clipboard hijack guard.
🧱 Network defense	A threat-intel firewall and private web/phishing blocking, driven from your OS's own firewall.
📦 Real installers	Native installers for Windows, macOS and Linux (x86_64 + ARM) - no bloatware, no upsells, hash-verifiable.

vs ClamAV: AetherAV adds an on-device LLM, behavioral + anti-rootkit + anti-ransomware engines, and cryptographically signed updates - things the classic open-source AV doesn't have.

---

🔬 How a scan works

Every file flows through independent engines. An exact hit short-circuits instantly; otherwise the layers vote and the worst verdict wins.

flowchart TD
    F([📄 File / bytes]) --> H{1 · Hash match?<br/>1.6M+ MD5/SHA1/SHA256}
    H -- yes --> BAD([☠ MALICIOUS])
    H -- no --> Y[2 · YARA-X rules · our own]
    Y --> N[3 · ClamAV patterns · 94k+]
    N --> P[4 · PE/ELF/Mach-O/PDF/Office/script parsers]
    P --> HE[5 · Static heuristics · entropy, W^X, packers]
    HE --> ML[6 · Static ML classifier]
    ML --> AI[7 · 🧠 Aegis-50M on-device LLM]
    AI --> SB[8 · Sandbox / shellcode emulation]
    SB --> IO[9 · Threat-intel IOC match · URLs/IPs/domains]
    IO --> REP{Known-good?<br/>cloud reputation · opt-in}
    REP -- yes --> CLEAN([✅ CLEAN])
    REP -- no --> V{Any layer<br/>flagged it?}
    V -- yes --> BAD
    V -- no --> CLEAN

    classDef file fill:#0e2731,stroke:#27d8ee,color:#d6e6ec
    classDef eng fill:#102a36,stroke:#2a5563,color:#cfe2e8
    classDef dec fill:#3a2e12,stroke:#f3a338,color:#ffe6b0
    classDef bad fill:#ef5350,stroke:#7a2230,color:#fff
    classDef good fill:#34e08a,stroke:#1b5e20,color:#04261a
    class F file
    class Y,N,P,HE,ML,AI,SB,IO eng
    class H,REP,V dec
    class BAD bad
    class CLEAN good

Loading

We don't publish self-reported detection percentages - they're meaningless unless you can reproduce them. Run the numbers yourself against a public corpus: ./tools/benchmark.sh --malware <dir> --clean <dir> (see docs/BENCHMARK.md). Independent results (VirusTotal contributor, then AV-Comparatives) are on the roadmap; until then, trust the method, not a marketing figure.

---

🛰️ Real-time protection

Beyond scanning files, AetherAV watches the live system and can block threats before they run.

flowchart LR
    subgraph LIVE [Live system]
      OA[📂 file open / exec]
      PR[⚙️ new process]
      MEM[🧠 process memory]
      NET[🌐 connections]
      FS[🗂️ your documents]
    end

    OA --> G1[on-access guard<br/>fanotify · allow/deny]
    PR --> G2[Process Sentinel<br/>new app? hidden? stealth?]
    MEM --> G3[memory scan<br/>injected / fileless code]
    NET --> G4[network monitor<br/>malware ports · bad IPs/domains]
    FS --> G5[ransomware shield<br/>canaries + rollback]

    G1 & G2 & G3 & G4 & G5 --> ENG[(AetherAV engine)]
    ENG --> ACT[🚫 block · 🔒 quarantine · rollback · 🔔 alert]

    classDef src fill:#102a36,stroke:#2a5563,color:#cfe2e8
    classDef grd fill:#0e2731,stroke:#27d8ee,color:#d6e6ec
    classDef eng fill:#173a2b,stroke:#34e08a,color:#d6f5e6
    classDef act fill:#3a1418,stroke:#ef5350,color:#ffd9d9
    class OA,PR,MEM,NET,FS src
    class G1,G2,G3,G4,G5 grd
    class ENG eng
    class ACT act

Loading

• Process Sentinel - flags any new app by its hash (not its name, which malware spoofs) and detects hidden processes via cross-view (a userland rootkit can't hide from it). Optional kernel-sourced live exec stream (cn_proc).
• Ransomware shield - plants canaries, detects mass-encryption, and rolls your files back from a snapshot.
• Memory scan - catches fileless / injected code (W^X, unbacked exec maps), and harvested wallet keys / seed phrases / cloud creds (stealer scraping).
• Threat firewall - blocks connections to malicious IPs / RAT-C2 ports via your OS's own firewall (nftables / netsh / pf), driven by our signed intel.
• Web / phishing protection - sinkholes malicious domains locally in the hosts file. No cloud, fully private (the big suites send every URL to theirs).
• Stealer & wallet shield - plants decoy wallet/credential files; any read is a near-certain infostealer, blocked at the source.
• Clipboard guard - detects crypto-address hijackers ("clippers") and restores your original address.
• Anti-exploit - flags exploit staging (NOP sleds, heap spray, stack-pivot gadgets, egg-hunters) in payloads and documents.

---

🔏 Updates you can't be tricked into trusting

The #1 risk for any AV is a poisoned update channel. AetherAV signs every feed and release with an Ed25519 key kept offline. Your client carries only the public key. A compromised server, CDN, or man-in-the-middle cannot forge an update - the client simply rejects it.

flowchart LR
    K[🔑 private key<br/>OFFLINE / air-gapped] -->|sign| FEED[📦 signed feed / release]
    FEED --> CDN[(🌍 any mirror / CDN<br/>untrusted)]
    CDN --> C[💻 your client]
    C --> CHK{verify with<br/>embedded public key}
    CHK -- valid --> OK([✅ apply update])
    CHK -- forged / tampered --> NO([🛑 rejected])

    ATK[😈 attacker hacks the server] -.->|can't sign| CDN

    classDef key fill:#1d2b3a,stroke:#9b8cff,color:#e7e1ff
    classDef obj fill:#102a36,stroke:#2a5563,color:#cfe2e8
    classDef dec fill:#3a2e12,stroke:#f3a338,color:#ffe6b0
    classDef good fill:#34e08a,stroke:#1b5e20,color:#04261a
    classDef bad fill:#ef5350,stroke:#7a2230,color:#fff
    class K key
    class FEED,CDN,C obj
    class CHK dec
    class OK good
    class NO,ATK bad

Loading

---

🤝 Community protection - without giving up your privacy

Sharing new malware makes everyone safer, but it must not expose you. So it's opt-in, anonymous, and send-only - and the analysis server never executes what it receives.

flowchart LR
    U[💻 user · opt-in] -->|hashes/IOCs<br/>sample only with consent| ING[🚪 ingestion<br/>never executes · anonymized]
    ING --> Q[(queue)]
    Q --> AN[🧪 analysis worker<br/>isolated · NO network]
    AN --> SIG[📝 new signatures]
    SIG -->|🔏 signed offline| FEED[📦 signed feed]
    FEED --> ALL[🌐 every user gets better detection]

    classDef u fill:#0e2731,stroke:#27d8ee,color:#d6e6ec
    classDef srv fill:#102a36,stroke:#2a5563,color:#cfe2e8
    classDef iso fill:#3a1418,stroke:#ef5350,color:#ffd9d9
    classDef good fill:#34e08a,stroke:#1b5e20,color:#04261a
    class U u
    class ING,Q,SIG,FEED srv
    class AN iso
    class ALL good

Loading

By default nothing leaves your machine. See server/README.md.

---

📦 Install

Real installers for every OS and CPU architecture - with a license step, component selection, and real-time protection set up for you. No bloatware, no upsells. Grab the one for your system from the latest release.

OS	Installer(s)	Architectures	Real-time
Windows	AetherAV-Setup-*.exe wizard (license · components · PATH · shortcuts · context menu)	x86_64 · arm64	on-access watcher auto-started (kernel minifilter scaffolded)
macOS	.pkg wizard, or a drag-to-Applications .dmg	universal (Intel + Apple Silicon)	LaunchDaemon
Linux	.deb package, or the install.sh wizard / .tar.gz	x86_64 · aarch64	systemd service (fanotify) + hourly signed updates

⚠️ Read this before installing - it's unsigned on purpose

AetherAV is not signed with a paid OS code-signing certificate. That's a deliberate choice, not a shortcut. A paid certificate only proves "someone paid a Certificate Authority and verified their identity" - it says nothing about whether the code is safe, and it can't be audited. Instead, AetherAV gives you a stronger, free trust path: reproducible builds + an Ed25519-signed SHA256SUMS. You (or anyone) can prove the binary came from this exact source - something a paid-signed black box can't offer. See docs/VERIFY.md.

The only cost of skipping the paid cert is that your OS will show a warning and installation needs administrator / root (it installs a real-time security service - any AV does). Here's exactly how to proceed on each system:

• 🪟 Windows - double-click AetherAV-Setup-*.exe. SmartScreen may say "Windows protected your PC" → click More info → Run anyway. Approve the UAC / "Run as administrator" prompt (needed to install the real-time service).
• 🍎 macOS - the .pkg/.dmg is from an "unidentified developer". Right-click the file → Open → Open (don't double-click), or run xattr -dr com.apple.quarantine <file>. Installing the system service asks for your password.
• 🐧 Linux - install and real-time protection need sudo: sudo apt install ./aetherav_*_*.deb (or sudo ./install.sh). The CLI gets the capabilities it needs for fanotify/firewall; full on-access still uses sudo.

Always verify what you downloaded (takes 5 seconds and beats any cert):

sha256sum -c SHA256SUMS          # bytes match the published hashes
aether verifyfile SHA256SUMS     # the hash list is Ed25519-signed by us -> ✓ TRUSTED

Prefer to build it yourself? Every installer is reproducible from source: installer/windows/build.ps1 · installer/macos/build-pkg.sh · installer/linux/build-deb.sh (or installer/linux/install.sh). A tagged release builds all of them for every architecture via GitHub Actions - see docs/RELEASE.md.

---

🚀 Quick start (from source)

# build
cargo build --release          # binary: target/release/aether

# scan a file or folder
aether scan ./Downloads -r

# real-time + advanced (Linux; some need root)
aether sentinel --learn        # learn a clean baseline of running apps
aether sentinel                # detect NEW / hidden / stealth processes
aether memscan                 # scan process memory for injected/fileless code
aether ransomguard ~/Documents # ransomware shield with rollback
sudo aether protect /          # kernel on-access blocking (fanotify)
aether netscan                 # live connections vs malware-port DB

# anti-theft & exploit defense
aether memscan --secrets                 # flag harvested wallet keys/seeds in memory
sudo aether stealerguard ~ --arm --watch --kill   # decoy traps + block infostealers
aether clipguard --watch --restore       # crypto clipboard-hijack guard
aether exploitscan suspicious.bin        # exploit-staging indicators (sleds, ROP, ...)

# network defense (prints rules; --apply installs into the OS firewall, needs admin)
sudo aether firewall --apply             # block malicious IPs/ports (nftables/netsh/pf)
sudo aether webprotect --apply           # block phishing/malware domains (hosts file)

# threat intel & reputation
aether intel ...               # manage feeds
aether reputation <file|hash>  # cloud known-good lookup (CIRCL, no key)
aether cve CVE-2021-44228      # vulnerability lookup
aether vt-scan <file>          # VirusTotal-contributor scan format

# run the engine as a service, or the desktop app
aether serve                   # HTTP/JSON API
./desktop/run.sh               # cross-platform GUI dashboard

Exit codes follow the ClamAV convention (0 clean, 1 threat found).

---

✅ Verify your download (don't just trust us)

sha256sum -c SHA256SUMS        # binary matches the published hash
aether verifyfile SHA256SUMS   # the hash list is signed by our offline key -> ✓ TRUSTED

Or rebuild from source and compare: ./tools/reproducible-build.sh. Full details: docs/VERIFY.md.

---

🧱 Architecture

A Rust workspace of focused crates; the engine is the core that every frontend (CLI, daemon, desktop) builds on.

flowchart TD
    subgraph FRONT [Frontends]
      CLI[aether CLI]
      DAEMON[HTTP/JSON daemon]
      GUI[Tauri desktop]
    end
    subgraph CORE [aether-core · orchestrator]
      ENGINES[hash · YARA · patterns · heuristics · ML · LLM · sandbox · anomaly · intel · cache]
    end
    subgraph SUP [Supporting crates]
      SIG[aether-signatures]
      PAR[aether-parsers]
      BEH[aether-behavior]
      RT[aether-realtime]
      FW[aether-firewall]
      INTEL[aether-intel]
      QUAR[aether-quarantine]
      LLM[aether-llm · Aegis-50M]
    end
    FRONT --> CORE --> SUP

    classDef f fill:#0e2731,stroke:#27d8ee,color:#d6e6ec
    classDef c fill:#173a2b,stroke:#34e08a,color:#d6f5e6
    classDef s fill:#102a36,stroke:#2a5563,color:#cfe2e8
    class CLI,DAEMON,GUI f
    class ENGINES c
    class SIG,PAR,BEH,RT,FW,INTEL,QUAR,LLM s

Loading

Detection content: ~1.64M file hashes · our own YARA-X rules · 94k+ pattern signatures · the Aegis-50M model · live abuse.ch / CIRCL intel.

We own our detection content. The default install ships only rules we wrote, plus databases we ingest and re-sign ourselves - so our protection never depends on a third-party rule server that could be compromised. An optional community YARA pack exists (tools/import_yara.py -> assets/ community-rules/) but is not loaded by default; review it before trusting it. YARA rules are data, not code, so a poisoned rule could at worst cause false positives - never run code on your machine.

---

🔒 Open source and hard to bypass

Being open is a trust advantage for an AV, not a weakness. Our defenses don't depend on secret code:

• Cryptographic integrity - signed updates & releases (forgery needs the offline key, not the source).
• Behavioral detection - malware can't hide what it does at runtime by reading our rules (encryption, injection, hidden processes, C2).
• Defense in depth - 10 layers; evading one isn't enough.
• Constantly-updated content - a static evasion goes stale fast.

Read the full threat model and disclosure policy in SECURITY.md.

---

🗺️ Status & roadmap

Done: 10 detection engines · on-device AI (Aegis-50M) · real-time protection (on-access, sentinel, memory, ransomware rollback) · threat firewall + web protection (3 OS) · anti-stealer / anti-clipper / memory key-scraping · anti-exploit · Authenticode publisher detection · signed feed + model updates · anonymous submission pipeline · desktop GUI with system tray + widget · native installers for every OS + CPU arch (Windows / macOS / Linux, x86_64 + ARM) + CI · Ed25519-verifiable releases · VirusTotal-contributor package · marketing website (web/) · reproducible builds.

Stage: active development / early beta - a working engine, not a finished product (see the honest-status note up top).

Earning trust (the credibility roadmap):

1. Reproducible public benchmark against an open corpus (MOTIF / SOREL / BODMAS) so anyone can run our numbers - see docs/BENCHMARK.md.
2. VirusTotal contributor (the vt-scan package is ready) for an independent detection signal.
3. AV-Comparatives → AV-TEST - see docs/CERTIFICATION.md.

Also next: ship the Windows kernel minifilter + macOS EndpointSecurity · deploy the public feed server · grow detection content. We intentionally skip paid OS code-signing - trust comes from reproducible builds + Ed25519-signed releases, not from a Certificate Authority (see docs/VERIFY.md).

📜 License

AetherAV is open source under the Apache License 2.0 - see LICENSE.

_{Built to be trusted: read it, build it, verify it.}