malmal 0.5.0

Upload scanner for images, SVG, PDF and PostScript files
Documentation

malmal

CI License: MIT OR Apache-2.0

malmal checks uploaded images, SVG, PDF and PostScript files before your application processes them. It answers three questions that antivirus usually does not:

  • Is the file what it claims to be? The content is compared with the file name and Content-Type.
  • Does it carry active content? Scripts, actions, embedded files, shell commands.
  • Is it well formed? Trailing data, polyglots, decompression bombs, truncation.

A file that cannot be analyzed completely is never reported as clean. malmal can also rebuild an image, SVG or PDF from what it draws, so that nothing else from the upload is kept.

It ships as a Rust library, a command-line tool, a daemon (HTTP, Unix socket, ClamAV protocol, ICAP), a C library and a Python package.

Contents

Supported formats

Format Type name Scanned Rebuilt as
JPEG jpeg yes JPEG
PNG, APNG png yes PNG
GIF gif yes GIF
WebP webp yes WebP (lossy stays lossy)
BMP bmp yes PNG
TIFF, BigTIFF tiff yes PNG; TIFF if it has several pages
ICO ico yes ICO
JPEG XL jxl yes PNG
PSD, PSB psd yes PNG
HEIC heif yes no
AVIF avif yes PNG
SVG svg yes SVG
PDF pdf yes PDF of page images (optional cdr-pdf feature)
PostScript, EPS, Illustrator (PS-based) postscript yes no

Any other file is reported as type.unknown. Office documents, archives and video are out of scope.

Installation

Prebuilt binaries for x86-64 and ARM64 Linux are attached to every release. They are static and need no other packages. Each archive contains malmal, malmald, libmalmal_ffi.a and malmal.h.

cargo binstall malmal-cli     # downloads the prebuilt binaries
cargo install malmal-cli      # builds from source

Container image:

docker run --rm -p 127.0.0.1:8080:8080 ghcr.io/nguyenthuan155632/malmal

Rust library:

cargo add malmal

Python package: wheels for x86-64 and ARM64 Linux are attached to each release. See bindings/python/README.md.

Ruby gems: malmal, with the library inside for Linux (glibc and musl) and macOS, and malmal-rails:

bundle add malmal-rails       # or bundle add malmal, outside Rails

From source:

sudo apt-get install musl-tools   # a C compiler for musl, for `ring` and, with YARA, `wasmtime`
CC_x86_64_unknown_linux_musl=musl-gcc \
  cargo build --release --workspace --target x86_64-unknown-linux-musl

The sandbox is available on Linux only. On macOS, build without --target, and do not pass --require-sandbox.

Quick start

malmal scan upload.pdf
upload.pdf: MALICIOUS (pdf)
  [High] pdf.javascript: document contains JavaScript at offset 186

The exit status gives the verdict, so a shell script can act on it directly:

malmal scan "$file" || rm -- "$file"
Exit status Meaning
0 Clean (for sanitize: the new file was written)
1 Suspicious
2 Malicious
3 Error: the file was not scanned

Treat anything other than 0 as a rejection.

Usage

Command line

malmal scan [OPTIONS] <FILE>...
malmal sanitize [OPTIONS] <INPUT> <OUTPUT>

scan checks each file; - reads standard input. With several files, the exit status is that of the worst one. --json prints one JSON object per file:

malmal scan --json incoming/*
{"path":"incoming/a.pdf","report":{"verdict":"malicious","detected_type":"pdf","declared_types":["pdf"],
 "size":285,"sha256":"8e38c368...","findings":[
  {"code":"pdf.javascript","severity":"high","message":"document contains JavaScript","offset":186}]}}

sanitize scans INPUT and, if it is clean, writes a rebuilt copy to OUTPUT (see Rebuilding files). OUTPUT is replaced atomically on success and left untouched otherwise.

Every finding has a stable code. Match on the code, not the message. All codes are listed in docs/findings.md.

See Configuration for all options.

Rust

use malmal::{ScanHints, Scanner, Verdict};

// Build once and share between threads: Scanner is Send + Sync.
let scanner = Scanner::builder().build();

let report = scanner.scan_bytes(&data, ScanHints {
    filename: Some("logo.svg"),           // as sent by the client
    content_type: Some("image/svg+xml"),
});

if report.verdict != Verdict::Clean {
    for finding in &report.findings {
        log::warn!("{} [{:?}] {}", finding.code, finding.severity, finding.message);
    }
    return Err(Rejected);
}
  • scan_bytes never fails or panics; anything that could not be analyzed becomes a finding.
  • scan_reader takes any impl Read and stops reading at the size limit.
  • A scan is CPU-bound and can run up to the time limit. In async code, run it on a blocking thread (tokio::task::spawn_blocking).
  • The time limit is checked between steps. scan_within(Arc<[u8]>, hints, wait) answers within wait even if a step never ends, with limit.time; see docs/integration.md.

A configured scanner:

use std::time::Duration;
use malmal::{FileType, HashList, Limits, Policy, Scanner};

let policy = Policy::default()
    .with_allowed_types([FileType::Jpeg, FileType::Png, FileType::Pdf])
    .allow("pdf.xfa")?;

let scanner = Scanner::builder()
    .policy(policy)
    .limits(Limits {
        max_file_size: 20 << 20,
        max_duration: Duration::from_secs(5),
        ..Limits::default()
    })
    .hashlist(HashList::parse(&std::fs::read_to_string("blocklist.txt")?)?)
    .pdf_passwords(["statement-password"])
    .build();

Complete axum and actix-web handlers are in docs/integration.md.

Cargo feature Adds
cdr Scanner::sanitize and malmal::cdr: rebuilding raster images
cdr-svg Rebuilding SVG (implies cdr)
cdr-pdf Rebuilding PDF as page images (implies cdr)
yara ScannerBuilder::yara and YaraRules

All features are off by default in the library.

Python

The package loads the C library with ctypes and has no other dependencies.

import malmal

scanner = malmal.Scanner()      # build once, share between threads
report = scanner.scan(data, filename="a.png", content_type="image/png")
if not report.clean:
    reject([finding["code"] for finding in report.findings])
scanner = malmal.Scanner(
    strict=False,                                # strict policy
    hashlist=open("blocklist.txt").read(),       # SHA-256 blocklist
    pdf_passwords=["statement-password"],        # str or bytes
    allow=["pdf.xfa", "pdf.embedded_file"],      # or "pdf.xfa|pdf.embedded_file"
    library=None,                                # path to libmalmal_ffi, if not found
)
Call Returns Raises
Scanner(...) a scanner ValueError for an invalid blocklist, password or code
scan(data, filename=None, content_type=None) Report ScanError if the file was not scanned
sanitize(data, filename=None, content_type=None, jpeg_quality=90) Sanitized Rejected (with .report), NotRebuilt, ScanError

Report has verdict, clean, detected_type, declared_types, size, sha256 and findings. Sanitized has data, file_type, width, height, frames and report. A Scanner is also a context manager.

Ruby and Rails

The malmal gem calls the same library through ffi, and releases the GVL while it scans. One scanner is shared by every thread:

SCANNER = Malmal::Scanner.new(policy: :documents, types: %i[jpeg png pdf], wait: 30)
report = SCANNER.scan(upload, filename: "a.pdf", content_type: "application/pdf")
reject(report.codes) unless report.clean?

In Rails, malmal-rails validates Active Storage attachments before they are stored, rebuilds the ones you name, and scans every stored blob:

class Document < ApplicationRecord
  has_one_attached :file
  validates :file, malmal: true                         # :malmal_malicious, with the codes
end

class User < ApplicationRecord
  has_one_attached :avatar
  sanitizes_attachment :avatar, types: %i[jpeg png webp]  # only the rebuilt file is kept
end

Both read config.malmal, which can also point at a running malmald. See bindings/ruby/malmal, bindings/ruby/malmal-rails, and the application in examples/rails.

C

Include malmal.h and link libmalmal_ffi.a. A scanner can be used from several threads at once.

#include <malmal.h>

char *error = NULL;
malmal_scanner *scanner = malmal_scanner_new_with_allow(
    0,                 /* strict policy: 0 or 1 */
    NULL, 0,           /* SHA-256 blocklist text, or NULL */
    NULL, 0,           /* PDF passwords, one per line, or NULL */
    "pdf.xfa",         /* codes to allow, or NULL */
    &error);
if (!scanner) {
    fprintf(stderr, "malmal: %s\n", error);
    malmal_string_free(error);
    return 1;
}

char *report = NULL;
int verdict = malmal_scan(scanner, data, len, "a.png", "image/png", &report);
if (verdict != MALMAL_CLEAN) {
    /* reject; `report` holds the JSON report, or NULL on MALMAL_ERROR */
}
malmal_string_free(report);
malmal_scanner_free(scanner);
Function Purpose
malmal_scanner_new(strict, hashlist, len) Create a scanner
malmal_scanner_new_with_passwords(..., pdf_passwords, len) Also try passwords on encrypted PDFs
malmal_scanner_new_with_allow(..., allow, &error) Also allow finding codes; reports why creation failed
malmal_scan(scanner, data, len, filename, content_type, &json) Scan a buffer
malmal_sanitize(scanner, data, len, filename, content_type, quality, &out, &out_len, &json) Scan and rebuild
malmal_string_free, malmal_bytes_free, malmal_scanner_free Release what the library returned

Return values are MALMAL_CLEAN, MALMAL_SUSPICIOUS, MALMAL_MALICIOUS, MALMAL_ERROR and, from malmal_sanitize only, MALMAL_NOT_REBUILT. The full contract is in malmal.h; an example is in examples/scan.c.

Daemon

malmald serves the scanner to any language over one or more endpoints:

malmald --http 127.0.0.1:8080 --socket /run/malmal.sock --require-sandbox
Option Protocol Use it for
--http <ADDR> HTTP Any language with an HTTP client
--socket <PATH> JSON over a Unix socket Local services; no HTTP overhead
--clamd <ADDR> ClamAV clamd protocol Software that already calls ClamAV
--icap <ADDR> ICAP Proxies such as Squid
--https, --clamd-tls, --icaps <ADDR> The same three over TLS Endpoints other hosts reach

ADDR is host:port or the path of a Unix socket (a TLS endpoint takes host:port only). Bind the plain endpoints to loopback or a Unix socket, and let only your upload service reach them.

  • --token-file <FILE> makes every HTTP request but GET /healthz, and every ICAP REQMOD and RESPMOD, carry Authorization: Bearer <token>, the first line of FILE (16 characters or more); without it the answer is 401. Squid sends it with adaptation_meta Authorization "Bearer <token>".
  • --tls-cert <FILE> and --tls-key <FILE> (PEM) serve the TLS endpoints; --tls-client-ca <FILE> also makes their clients present a certificate signed by one of the CAs in FILE. That is the way to authenticate the ClamAV protocol, which has no place for a token, from another host; on the same host, a Unix socket (mode 0660) is.

At start-up the daemon warns about every endpoint that is not on loopback and that anyone who reaches it can use.

HTTP. The request body is the file itself, not a multipart form. Any status other than 200 means the file was not scanned.

curl --data-binary @upload.pdf -H 'Content-Type: application/pdf' \
  'http://127.0.0.1:8080/scan?filename=upload.pdf'
Request Response
POST /scan 200 and {"ok": true, "report": {...}}, whatever the verdict
POST /sanitize 200 and the rebuilt file; 422 if not clean, 415 if it cannot be rebuilt
GET /healthz 200 ok
GET /metrics Prometheus counters: scans by verdict, errors, bytes, time, rebuilt files

The file name can also be sent in an X-Filename header.

Unix socket. Each request is one JSON line followed by size bytes of file; the reply is one JSON line. A connection can carry many requests.

import json, socket

def scan(data, name):
    with socket.socket(socket.AF_UNIX) as sock:
        sock.connect("/run/malmal.sock")
        sock.sendall(json.dumps({"size": len(data), "filename": name}).encode() + b"\n" + data)
        reply = json.loads(sock.makefile().readline())
    if not reply.get("ok"):
        raise RuntimeError(reply.get("error", "not scanned"))
    return reply["report"]

ClamAV protocol. PING, VERSION and INSTREAM are supported. The protocol has only two answers, so suspicious files are reported as FOUND, like malicious ones (stream: Malmal.Suspicious.pdf.encrypted FOUND). It carries no file name, so the name and type are not checked.

ICAP. With Squid:

icap_enable on
icap_service malmal_req reqmod_precache icap://127.0.0.1:1344/scan bypass=off
icap_service malmal_resp respmod_precache icap://127.0.0.1:1344/scan bypass=off
acl uploads method POST PUT
acl downloads urlpath_regex -i \.(png|jpe?g|gif|webp|pdf)$
adaptation_access malmal_req allow uploads
adaptation_access malmal_resp allow downloads

bypass=off makes Squid refuse requests when the daemon is unreachable. Each part of a multipart/form-data upload is scanned separately; a message that is not clean is replaced by 403 Forbidden. Send only uploads and downloads of files, as above: an HTML page or a JSON request is not a supported file type and would be blocked.

Status codes, metrics and protocol details are in docs/integration.md.

Configuration

Options

malmal and malmald accept the same scanning options.

Option Default Rust Python C
--policy default|strict|documents default Policy::strict(), Policy::documents() strict=True strict
--token-file <FILE> (malmald) none
--tls-cert, --tls-key, --tls-client-ca <FILE> (malmald) none
--types <LIST> all types Policy::with_allowed_types
--allow <CODES> none Policy::allow allow= allow
--hashlist <FILE> none ScannerBuilder::hashlist hashlist= hashlist
--pdf-passwords <FILE> none ScannerBuilder::pdf_passwords pdf_passwords= pdf_passwords
--no-hash off ScannerBuilder::without_hash
--max-size <BYTES> 64 MiB Limits::max_file_size
--max-decompressed <BYTES> 256 MiB Limits::max_decompressed
--max-pixels <N> 2^28 Limits::max_pixels
--max-memory <BYTES> 512 MiB ScannerBuilder::memory_limit max_memory= malmal_scanner_set_max_memory
--scan-timeout <S> 10 Limits::max_duration
--hard-timeout <S> 6 × scan timeout
--yara <PATH> none ScannerBuilder::yara
--require-sandbox off
--no-sandbox off

documents is the default policy with forms (pdf.xfa) and attachments (pdf.embedded_file) allowed; what is found inside them still counts. From Python and C, pass those two codes to allow.

The Python and C interfaces use the default limits. Embedded files are scanned up to Limits::max_depth (4) levels deep.

malmal sanitize also takes:

Option Default Effect
--jpeg-quality <N> 90 Quality of rebuilt JPEG files, 1–100
--keep-cmyk off Keep the inks of a CMYK JPEG, TIFF or PSD, with their profile, rather than turn them into sRGB
--keep-profile off Keep the colors of an image under an RGB or grey profile other than sRGB (or the primaries an AVIF names), with the profile rewritten, rather than turn them into sRGB
--pdf-dpi <N> 150 Resolution of rebuilt PDF pages, 36–600
--pdf-jpeg off Store a PDF page as JPEG when that is smaller
--pdf-no-text off Leave out the invisible text layer of a rebuilt PDF
--pdf-raster-text off Keep all text of a rebuilt PDF in the page images

malmald also takes:

Option Default Effect
--workers <N> 4 Files scanned at once; bounds CPU and the memory scans take
--connections <N> 256 Connections kept open at once
--file-budget <BYTES> workers × max size Most that the files held at once may come to, as they arrive and while they are scanned
--timeout <S> 30 Idle time before a connection is closed
--min-rate <BYTES> 1024 Bytes per second a request must arrive at once --timeout has passed; 0 for any rate
--spool-dir <DIR> none Keep files of 1 MiB or more in temporary files in DIR instead of memory

Severity and policy

Each finding has a severity. The policy turns the most severe finding into the verdict:

Policy Info Low Medium High, Critical
default clean suspicious suspicious malicious
strict clean suspicious malicious malicious

We recommend rejecting every file that is not clean. The two policies then reject the same files; strict only changes which are labelled malicious.

Allowing finding codes

If your users legitimately send files with a feature malmal flags, allow that finding code instead of ignoring the verdict:

malmal scan --allow 'pdf.xfa|pdf.embedded_file' tax-form.pdf

Codes are separated by , or |, and --allow can be repeated. An allowed finding is still reported, at info, and the rest of the file is still checked: attachments are still scanned, and a script inside an XFA form is still reported as pdf.javascript.

An unknown code is an error, so a typo cannot silently allow nothing. limit.* codes cannot be allowed, because a file that was not fully analyzed is never clean; raise the limit instead.

Codes that legitimate files often carry:

Code Severity Common in Risk if allowed
pdf.xfa Medium Tax and government forms Forms render differently between readers
pdf.embedded_file Medium Source documents, e-invoices (ZUGFeRD, Factur-X); not C2PA manifests, which are pdf.c2pa_manifest at Info Attached files the reader can open; each is still scanned
pdf.javascript High Forms that calculate or validate fields Scripts run by the reader; the most common PDF attack
pdf.open_action, pdf.additional_action Medium/High Documents that open at a page or react to events Actions that run without a click
pdf.rich_media Medium Brochures with video, sound or 3D Media decoders with a history of exploits
pdf.external_ref Medium Prepress files, documents that include other files Rendered content that was not scanned
pdf.encrypted Medium Documents locked with a password you do not have Content that was not scanned at all
pdf.submit_form, pdf.remote_goto Low Forms that submit data, links to other documents Requests to addresses the file chooses
pdf.trailing_data, image.trailing_data Low Files saved by some editors and cameras; the trailers of Samsung and OPPO cameras and motion photos are image.vendor_trailer at Info Unread bytes after the end of the file
pdf.link_lure Low Flyers and covers that are one large picture linking to a site A click anywhere on the page opens a site the file chooses
pdf.launch High Old documents that open a local file Starts a program on the reader's machine
svg.foreign_object Medium Diagrams from tools such as draw.io HTML rendered inside the image
svg.external_ref Medium Images that load web fonts or pictures Requests to other sites when displayed
type.mismatch Medium Files with the wrong extension Files served as the wrong type

Never allow codes that only attacks produce: type.executable, payload.*, hash.blocklisted, yara.match, *.polyglot, *.malformed, svg.script, svg.event_handler, svg.javascript_uri, svg.entity, pdf.obfuscated_name, pdf.dangerous_uri, pdf.postscript, pdf.xml_entity, and the ps.* codes for file and device access.

In Rust, Policy::with_override(code, severity) sets any severity, for example to reject something the default policy accepts:

use malmal::{Policy, Severity};

let policy = Policy::default()
    .with_override("pdf.risky_codec", Severity::Medium);  // refuse JBIG2 and JPEG 2000 images

File types

All supported types are accepted by default. To accept fewer, list them; other files are reported as type.not_allowed:

malmal scan --types jpeg,png,webp,pdf upload.eps

Encrypted PDFs

PDFs encrypted with an empty password are always decrypted and inspected. Others are reported as pdf.encrypted (Medium) unless you supply their password. The password file has one password per line; each is tried as user and owner password.

malmal scan --pdf-passwords passwords.txt statement.pdf

Blocklist and YARA

A SHA-256 blocklist rejects known files. The output of sha256sum can be used directly:

sha256sum known-bad/* > blocklist.txt
malmal scan --hashlist blocklist.txt upload.png

YARA rules are matched against every file and every file embedded in a PDF. They need a build with the yara feature, which makes the binary about 16 MB; for a static musl binary, build it as above with --features yara. Rules can import the hash, math, string, pe and elf modules:

cargo install malmal-cli --features yara
malmal scan --yara rules/ upload.jpg

--yara takes a rule file or a directory of .yar and .yara files, and can be repeated.

Rebuilding files

A scan reports what a scanner can see. Rebuilding goes further: the file is decoded and a new one is written from its content, so metadata, comments, unknown chunks and appended data are dropped.

malmal sanitize upload.jpg stored.jpg
curl --fail --data-binary @upload.jpg -o stored.jpg http://127.0.0.1:8080/sanitize
// malmal = { version = "0.2", features = ["cdr"] }
use malmal::cdr::{Error, Options};

match scanner.sanitize(&data, hints, &Options::default()) {
    Ok(sanitized) => store(&sanitized.image.data, sanitized.image.file_type),
    Err(Error::Rejected(report) | Error::Unclean(report)) => reject(&report.findings),
    Err(other) => reject_with(other.to_string()),  // unsupported, too large, undecodable
}
try:
    image = scanner.sanitize(data, filename="a.png")
except malmal.SanitizeError:        # Rejected or NotRebuilt
    reject()
else:
    store(image.data, image.file_type)

In C, use malmal_sanitize. In every interface, if rebuilding fails, store nothing.

  • Images keep their pixels, animation frames, frame delays and loop count. Colors are converted to sRGB and the color profile is dropped. Where it can, the picture is copied as it was coded rather than decoded and encoded again, so nothing of it is lost:

    • JPEG keeps its coefficients and is rotated as its Exif orientation says; one that cannot be rotated or converted that way is re-encoded at the chosen quality at most.
    • Lossy WebP keeps every symbol of its VP8 stream, coded again at the same probabilities, and is kept only if it decodes to exactly the same pixels; it comes out the size it went in. Its alpha is coded again without loss, and its Exif orientation is kept, as Exif of nothing else.
    • GIF keeps its palettes and pixel indices; APNG keeps its frames as they were laid out, at their own depth.
    • A TIFF of JPEG strips or tiles becomes one JPEG of their coefficients, or a TIFF of such pages if it has several (or inks kept with --keep-cmyk); a JPEG XL made from a JPEG gives back that JPEG. Each is checked against the pixels it shows. JPEG in a TIFF of the old kind (compression 6) becomes the JPEG it stands for.

    BMP, PSD (the composite image saved with the layers; a duotone in the colors of its inks, overprints included), JPEG XL and AVIF become PNG, a JPEG XL animation an animated PNG, and a TIFF a PNG of its one page (a TIFF if its samples are floating point) or a TIFF of all of them, so store the file under the type the result reports. Lossless WebP is written by an encoder of malmal's own and checked pixel for pixel. TIFF is read in every compression libtiff writes (CCITT fax, LZMA and Zstandard among them) and at any depth; AVIF is decoded by filmcraft-av1, safe Rust that matches libdav1d sample for sample, with its grid, alpha, crop, rotation and mirroring. An ICO keeps every image. A PNG of few colors stays indexed, and the pixel density is kept.

  • CMYK JPEG, TIFF and PSD are converted to sRGB through their profile. With --keep-cmyk (Inks::Keep, keep_cmyk in the bindings) they keep their inks instead: a JPEG its coefficients, a TIFF or PSD its samples, written as a TIFF. The profile is kept with them, rebuilt from only the tags that define its colors, copied byte for byte and checked to print every ink exactly as the original does.

  • Wide-gamut RGB (Display P3, Adobe RGB) images are converted to sRGB, which clips colors sRGB cannot show, and grey under a profile of its own to sRGB's curve. With --keep-profile (Profiles::Keep, keep_profile in the bindings) a JPEG, PNG, TIFF, WebP, BMP, GIF, animated PNG, JPEG XL, PSD or AVIF keeps its samples (a JPEG its coefficients, a lossy WebP its VP8 stream) and its profile, rebuilt the same way and checked on a grid of every color mix; an AVIF that names its primaries and transfer keeps them in the PNG's cICP chunk.

  • SVG is filtered through an allowlist (svg-hush) that keeps only what draws. The result is scanned again and refused if it is not clean.

  • PDF pages are rendered (hayro) and the new PDF contains the page images and the page's characters where they stood: text can still be selected, searched and copied (--pdf-no-text or Options::pdf_text = false leaves it out). Text in one opaque colour on a plain background is taken out of the image, once each glyph is checked against it, and drawn from the outlines of its glyphs, in fonts written anew, so it stays sharp at any size (--pdf-raster-text or Options::pdf_vector_text = false keeps it in the image); other text stays in the image, over characters that draw nothing. No font program, link or form is kept. On clean documents of CIC-Evasive-PDFMal2022, those with text come out about five times the size of the original (median), and scans about three times. This needs the cdr-pdf feature, which is not included in the released binaries, container or wheels:

    cargo install malmal-cli --features cdr-pdf
    

Every rebuilt file is scanned again, because pixels themselves can encode another file.

Deployment

Sandbox

Before reading a file, malmal and malmald restrict themselves: Landlock limits file access to the inputs, seccomp blocks execve, new sockets and other dangerous system calls, and an rlimit caps address space. Pass --require-sandbox in production so that the process exits if the sandbox cannot be applied. Linux only.

Limits

A file that exceeds a limit gets a limit.* finding and is never clean. A scan that runs past --hard-timeout ends the process with exit status 3, so run the daemon under a supervisor:

[Service]
ExecStart=/usr/local/bin/malmald --socket /run/malmal/malmal.sock --http 127.0.0.1:8080 --require-sandbox
Restart=always
DynamicUser=yes
RuntimeDirectory=malmal

Memory

malmal scan maps files instead of reading them, so memory depends on the structure of a file, not its size: a 1 GB PDF is scanned within a 96 MB memory limit. The daemon keeps uploads in memory unless --spool-dir is set; with it, files of 1 MiB or more are kept in unnamed temporary files that are removed when the request ends.

What the daemon holds is the files themselves, while they arrive and while they are scanned, and what the scans take beyond them. --file-budget bounds the first: a file takes its share as it arrives and waits for room when there is none, and a request that cannot get room within --timeout is refused (503 over HTTP), never answered as clean. The default, --workers times --max-size, is what the files could take before the option existed. The second is set per scan by --max-decompressed and --max-pixels, and in total by --workers, and --max-memory (512 MiB, 0 for none) stops a scan that holds more than that beyond what its thread held when it began, with limit.memory. Two files that each inflate to 200 MB peaked at 205 MB without it and 55 MB with a limit of 64 MiB; eight of them sent to a daemon with four workers, at 807 MB and 201 MB. The ceiling is then about --file-budget plus --workers times --max-memory; lower these options to fit your memory. Details and measurements are in docs/integration.md.

Container

The image runs malmald --http 0.0.0.0:8080 as an unprivileged user. Arguments replace the default command:

docker run --rm -p 127.0.0.1:3310:3310 ghcr.io/nguyenthuan155632/malmal \
  --clamd 0.0.0.0:3310 --max-size 20000000

To reach the HTTP endpoint from other hosts, give it a token (the file has to be readable by user 65534, whom the image runs as):

docker run --rm -p 8080:8080 -v /etc/malmal/token:/token:ro ghcr.io/nguyenthuan155632/malmal \
  --http 0.0.0.0:8080 --token-file /token

Accuracy and performance

Results on CIC-Evasive-PDFMal2022 (9,107 clean and 21,899 malicious PDF files):

Configuration Malicious rejected Clean accepted
Default policy 99.07% 87.32%
--policy documents 99.06% 92.38%
--allow 'pdf.xfa|pdf.embedded_file|pdf.javascript' 80.15% 97.11%

Every clean file that is rejected carries active content or an attachment: JavaScript (660), XFA forms (260), attachments (207), launch actions (12), media (9) and a few others. Of the 203 malicious files accepted, 187 contain a link to an ordinary web address and nothing else: one smaller than a fifth of the page (or pdf.link_lure would refuse it), that goes through no shortener and to no numeric, punycode or made-up name (or pdf.disguised_link would). Judging such an address requires a URL reputation service.

To measure false positives on your own files before you deploy:

cargo run --release --example survey -- --expect-clean DIR...

Scan times on a six-core desktop from 2018 (files in the page cache):

File Size Time With --no-hash Memory
PDF, mostly images 60 MB 0.26 s 0.03 s 4 MB
PDF, mostly images 1 GB 4.5 s 0.41 s 4 MB
PDF, 100,000 objects 10 MB 0.18 s 0.18 s 12 MB
PDF, 500,000 objects 51 MB 0.89 s 0.88 s 42 MB
PDF, 200 MB of page content once inflated 8 MB 0.63 s 0.63 s 5 MB

Memory is the peak resident size minus the mapped file. For large files most of the time goes to the SHA-256 hash; --no-hash skips it when you do not need it.

Limitations

  • PostScript is a programming language, so static analysis cannot be complete. Code built at run time is reported as ps.exec_dynamic, not analyzed. Always run Ghostscript with -dSAFER and in a sandbox.
  • PDF JavaScript is reported because it is present, not because of what it does.
  • Links in PDFs are not judged by where they go: a document whose only link is to a malicious site is clean, unless the link covers a fifth of a page (pdf.link_lure) or hides where it leads (pdf.disguised_link).
  • Pixel data is not decoded during a scan. For HEIC and AVIF the container is checked, not the HEVC or AV1 stream (rebuilding an AVIF does decode it).
  • Encrypted PDFs support the standard security handler only, not certificate encryption.
  • HEIC is scanned but not rebuilt (there is no HEVC decoder in Rust), nor is a PSD saved without a composite image ("Maximize compatibility" off). AVIF in a range for high dynamic range displays (PQ, HLG) becomes a 16-bit PNG that names its primaries and transfer (cICP): viewers that do not read that chunk show it wrong. An AVIF image sequence without a still image is refused.
  • Rebuilt images keep their pixels, and anything encoded in them. Rebuilt files are scanned again for that reason; if you call malmal::cdr::rebuild directly, scan the result yourself.
  • The library cannot stop a step that does not finish. The time limit is checked between steps. Scanner::scan_within answers in time anyway, with limit.time, but leaves the scan running on its thread (at most MAX_OVERDUE of them); only the command-line tools end such a scan, with --hard-timeout.
  • The ClamAV protocol has no token. From other hosts, authenticate it with --clamd-tls and --tls-client-ca. TLS certificates are read at start-up; a renewed certificate needs a restart.
  • Slow clients are let go, not refused. A request that arrives at --min-rate or faster keeps its connection and the part of --file-budget it has sent, however long it takes; enough such clients fill --connections or the budget, and the others wait. They hold no worker, so the files that have arrived are still scanned.
  • Scans do not wait for memory. --file-budget counts the files, not what scanning them takes; that is bounded per scan by --max-memory and in total by --workers. The memory limit is checked as PDF streams, PNG, SVGZ and JPEG XL data are inflated and between steps, so a step that allocates at once (a decoder of the image crate, say) can pass it before it is caught.
  • macOS has no sandbox, and the test suite is not run there.

Development

Building needs Rust 1.95 or later (rust-version); fuzzing needs nightly.

cargo test --workspace
cd fuzz && cargo +nightly fuzz run pdf corpus/pdf seeds/pdf
  • tests/corpus/clean and tests/corpus/malicious hold sample files. tests/corpus/expected.tsv records the expected findings for each; after an intended change, update it with MALMAL_RECORD=1 cargo test -p malmal --test corpus.
  • cargo run --release --example evaluate -- --clean DIR --malicious DIR measures detection and false positives on labelled files. --allow, --strict and --tsv are supported.
  • cargo run --release --example throughput -- FILE... measures throughput.
  • fuzz/ has one target per format; list them with cargo fuzz list.

See CHANGELOG.md for release notes.

License

Licensed under either of Apache License, Version 2.0 or MIT license, at your option.