Packj (pronounced package) is a tool to mitigate software supply chain attacks. It can detect malicious, vulnerable, abandoned, typo-squatting, and other "risky" packages from popular open-source package registries, such as NPM, RubyGems, and PyPI. It can be easily customized to minimize noise. Packj started as a PhD research project and is currently being developed under various govt grants.
Note Self-hosted Packj webserver and several integrations coming later this month π Watch this repo to stay up to date.
- Get started - available as Docker image, GitHub Action, and packages
- Functionality - deep static/dynamic code analysis and sandboxing
- Supported ecosystems - NPM, PyPI, Rubygems, PHP, Rust
- Our story - started as a PhD research project and is backed by govt grants
- Why Packj - existing CVE scanners ASSUME code is BENIGN and not analyze its behavior
- Customization - turn off alerts as per your threat model to reduce noise
- Malware found - reported over 70 malicious PyPI and RubyGems packages
- Talks and videos - presentations from PyCon, OpenSourceSummit, BlackHAT
- Project roadmap - view or suggest new features; join our discord channel
- Team and collaboration - lead by Cybersecurity researchers from academia/industry
- FAQ - supported package managers, commonly asked questions on techniques, and more
We support multiple deployment models:
Use Packj to audit dependencies in pull requests.
- name: Packj Security Audit
uses: ossillate-inc/[email protected]
with:
# TODO: replace with your dependency files in the repo
DEPENDENCY_FILES: pypi:requirements.txt,npm:package.json,rubygems:Gemfile
REPO_TOKEN: ${{ secrets.GITHUB_TOKEN }}
View on GitHub marketplace. Example PR run.
The quickest way to try/test Packj is using Docker. Podman is also supported for containerized (isolated) runs.
docker run -v /tmp:/tmp/packj -it ossillate/packj:latest --help
Clone this repo,
https://github.com/ossillate-inc/packj.git && cd packj
Install dependencies
bundle install && pip3 install -r requirements.txt
Start with help:
python3 main.py --help
Packj can vet pubished packages from NPM, PyPI, Rust, PHP, and Rubygems package registries. Rust and PHP support is WIP. We're actively adding support for registries. It also supports vetting local (unpublished) NPM and PyPI packages.
Packj offers the following tools:
Packj audits open-source software packages for "risky" attributes that make them vulnerable to supply chain attacks. For instance, packages with expired email domains (lacking 2FA), large release time gap, sensitive APIs or access permissions, etc. are flagged as risky.
Auditing the following is supported:
- multiple packages:
python3 main.py audit -p pypi:requests rubygems:overcommit
- dependency files:
python3 main.py audit -f npm:package.json pypi:requirements.txt
By default, audit
only performs static code analysis to detect risky code. You can paas -t
or --trace
flag to perform dynamic code analysis as well, which will install all requested packages under strace and monitor install-time behavior of packages. Please see the example output below.
$ docker run -v /tmp:/tmp/packj -it ossillate/packj:latest audit --trace -p npm:browserify
[+] Fetching 'browserify' from npm..........PASS [ver 17.0.0]
[+] Checking package description.........PASS [browser-side require() the node way]
[+] Checking release history.............PASS [484 version(s)]
[+] Checking version........................RISK [702 days old]
[+] Checking release time gap............PASS [68 days since last release]
[+] Checking author.........................PASS [[email protected]]
[+] Checking email/domain validity.......RISK [expired author email domain]
[+] Checking readme.........................PASS [26838 bytes]
[+] Checking homepage.......................PASS [https://github.com/browserify/browserify#readme]
[+] Checking downloads......................PASS [2M weekly]
[+] Checking repo URL.......................PASS [https://github.com/browserify/browserify]
[+] Checking repo data...................PASS [stars: 14189, forks: 1244]
[+] Checking if repo is a forked copy....PASS [original, not forked]
[+] Checking repo description............PASS [browser-side require() the node.js way]
[+] Checking repo activity...............PASS [commits: 2290, contributors: 207, tags: 413]
[+] Checking for CVEs.......................PASS [none found]
[+] Checking dependencies...................RISK [48 found]
[+] Downloading package from npm............PASS [163.83 KB]
[+] Analyzing code..........................RISK [needs 3 perm(s): decode,codegen,file]
[+] Checking files/funcs....................PASS [429 files (383 .js), 744 funcs, LoC: 9.7K]
[+] Installing package and tracing code.....PASS [found 5 process,1130 files,22 network syscalls]
=============================================
[+] 5 risk(s) found, package is undesirable!
=> Complete report: /tmp/packj_54rbjhgm/report_npm-browserify-17.0.0_hlr1rhcz.json
{
"undesirable": [
"old package: 702 days old",
"invalid or no author email: expired author email domain",
"generates new code at runtime",
"reads files and dirs",
"forks or exits OS processes",
]
}
WARNING: since packages could execute malicious code during installation, it is recommended to ONLY use
-t
or--trace
when running inside a Docker container or a Virtual Machine.
Audit can also be performed in Docker/Podman containers. Please find details on risky attributes and how to use at Audit README.
Packj offers a lightweight sandboxing for safe installation
of a package. Specifically, it prevents malicious packages from exfiltrating sensitive data, accessing sensitive files (e.g., SSH keys), and persisting malware.
It sandboxes install-time scripts, including any native compliation. It uses strace (i.e., NO VM/Container required).
Please find details on the sandboxing mechanism and how to use at Sandbox README.
$ python3 main.py sandbox gem install overcommit
Fetching: overcommit-0.59.1.gem (100%)
Install hooks by running `overcommit --install` in your Git repository
Successfully installed overcommit-0.59.1
Parsing documentation for overcommit-0.59.1
Installing ri documentation for overcommit-0.59.1
#############################
# Review summarized activity
#############################
[+] Network connections
[+] DNS (1 IPv4 addresses) at port 53 [rule: ALLOW]
[+] rubygems.org (4 IPv6 addresses) at port 443 [rule: IPv6 rules not supported]
[+] rubygems.org (4 IPv4 addresses) at port 443 [rule: ALLOW]
[+] Filesystem changes
/
βββ home
βββ ubuntu
βββ .ruby
βββ gems
β βββ iniparse-1.5.0 [new: DIR, 15 files, 46.6K bytes]
β βββ rexml-3.2.5 [new: DIR, 77 files, 455.6K bytes]
β βββ overcommit-0.59.1 [new: DIR, 252 files, 432.7K bytes]
β βββ childprocess-4.1.0 [new: DIR, 57 files, 141.2K bytes]
βββ cache
β βββ iniparse-1.5.0.gem [new: FILE, 16.4K bytes]
β βββ rexml-3.2.5.gem [new: FILE, 93.2K bytes]
β βββ childprocess-4.1.0.gem [new: FILE, 34.3K bytes]
β βββ overcommit-0.59.1.gem [new: FILE, 84K bytes]
βββ specifications
β βββ rexml-3.2.5.gemspec [new: FILE, 2.7K bytes]
β βββ overcommit-0.59.1.gemspec [new: FILE, 1.7K bytes]
β βββ childprocess-4.1.0.gemspec [new: FILE, 1.8K bytes]
β βββ iniparse-1.5.0.gemspec [new: FILE, 1.3K bytes]
βββ bin
β βββ overcommit [new: FILE, 622 bytes]
βββ doc
βββ iniparse-1.5.0
β βββ ri [new: DIR, 119 files, 131.7K bytes]
βββ rexml-3.2.5
β βββ ri [new: DIR, 836 files, 841K bytes]
βββ overcommit-0.59.1
β βββ ri [new: DIR, 1046 files, 1.5M bytes]
βββ childprocess-4.1.0
βββ ri [new: DIR, 272 files, 297.8K bytes]
[C]ommit all changes, [Q|q]uit & discard changes, [L|l]ist details:
TL;DR Packj started as a PhD research project. It is backed by various government grants.
Packj started as an academic research project. Specifically, the static code analysis techniques used by Packj are based on cutting-edge Cybersecurity research: MalOSS project by our research group at Georgia Tech.
Packj is backed by generous grants from NSF, GRA, and ALInnovate.
TL;DR The state-of-the-art vulnerability scanners assume that the third-party open-source code is BENIGN. Therefore, all such tools ONLY address threats from accidental programming bugs in benign code (a.k.a. CVEs such as Log4J). They DO NOT protect against Solarwinds-like modern software supply-chain attacks from deliberately bad (a.k.a. malicious) code that is propagated by bad actors using new vulnerabilities in the supply channel, including dependency confusion, typo-squatting, protestware (sabotaging), account hijacking, and social engineering. A recent (Dec'22) example is PyTorch package that was compromised using dependency confusion vulnerability (no CVE assigned).
Packj not only audits for CVEs, but also performs deep static+dynamic code analysis as well as metadata checks to detect any "risky" behavior and attributes, such as spawning of shell, use of SSH keys, mismatch of GitHub code vs packaged code (provenance), lack of 2FA, and several more. Such insecure attributes do not qualify as as CVEs, which is why none of the existing tools can flag them. Packj can flag malicious, typo-squatting, abandoned, vulnerable, and other insecure dependencies (weak links) in your software supply chain.
The current software supply chain threat model assumes that the third-party open-source code is benign, and therefore, security vulnerabilities are only tracked for accidental programming bugs (a.k.a. CVEs). As such, all existing open-source vulnerability scanners ONLY report publicly known CVEs and address threats from accidental bugs in benign code.
A typical example of an accidental programming bug is a missing bounds check on user input, which makes the code vulnerable to buffer overflow attacks. Real-world popular examples include Log4J and HeartBleed. Attackers need to develop an exploit to trigger CVEs (e.g., a crafted TCP/IP packet in case of HeartBleed or a numerically high input to cause buffer overflow). CVEs can be fixed by patching or upgrading to a newer version of the library (e.g., newer version of Log4J fixes the CVE).
The modern software supply chain threat landscape shifted after the Solarwinds attack. Bad actors have found new vulnerabilities, but this time in the supply channel, not code. These new vulnerabilities such as dependency confusion, typo-squatting, protestware (sabotaging), account hijacking, and social engineering are being exploited to propagate malware. Thousands of compromised NPM/PyPI/Ruby packages have been reported.
In contrast to CVEs, malware is deliberately bad (a.k.a. malicious) code. Moreover, malware itself is an exploit and cannot be patched or fixed by upgrading to a newer version. For example, dependency confusion attack was intentionally malicious; it did not exploit any accidental programming bug in the code. Similarly, an author of popular package sabotaging their own code to protest against the war is very much intentional and does not exploit any CVEs. Typo-squatting is another attack vector that bad actors use to propagate malware in popular open-source package registries: it exploits typos and inexperience of devs, not accidental programming bugs or CVEs in the code.
Existing scanners FAIL to detect these Solarwinds-like modern software supply-chain attacks from deliberately vulnerable (malicious) code. These tools simply scan the source code for open-source dependencies, compile a list of all dependencies being used, and look each <dependency-NAME, dependency-VERSION> up in a database (e.g., NVD) to report affected package versions (e.g., vulnerable version of Log4J, LibSSL version affected by HeartBleed).
Packj not only audits for CVEs, but also performs deep static+dynamic code analysis as well as metadata checks to detect any "risky" behavior and attributes, such as spawning of shell, use of SSH keys, mismatch of GitHub code vs packaged code (provenance), lack of 2FA, and several more. Such insecure attributes do not qualify as as CVEs, which is why none of the existing tools can flag them. Packj can flag malicious, typo-squatting, abandoned, vulnerable, and other insecure dependencies (weak links) in your software supply chain. Please read more at Audit README
Packj can be easily customized (zero noise) to your threat model. Simply add a .packj.yaml file in the top dir of your repo/project and reduce alert fatigue by commenting out unwanted attributes.
We found over 40 and 20 malicious packages on PyPI and Rubygems, respectively using this tool. A number of them been taken down. Refer to an example below:
$ python3 main.py audit pypi:krisqian
[+] Fetching 'krisqian' from pypi...OK [ver 0.0.7]
[+] Checking version...OK [256 days old]
[+] Checking release history...OK [7 version(s)]
[+] Checking release time gap...OK [1 days since last release]
[+] Checking author...OK [[email protected]]
[+] Checking email/domain validity...OK [[email protected]]
[+] Checking readme...ALERT [no readme]
[+] Checking homepage...OK [https://www.bilibili.com/bangumi/media/md140632]
[+] Checking downloads...OK [13 weekly]
[+] Checking repo_url URL...OK [None]
[+] Checking for CVEs...OK [none found]
[+] Checking dependencies...OK [none found]
[+] Downloading package 'KrisQian' (ver 0.0.7) from pypi...OK [1.94 KB]
[+] Analyzing code...ALERT [needs 3 perms: process,network,file]
[+] Checking files/funcs...OK [9 files (2 .py), 6 funcs, LoC: 184]
=============================================
[+] 6 risk(s) found, package is undesirable!
{
"undesirable": [
"no readme",
"only 45 weekly downloads",
"no source repo found",
"generates new code at runtime",
"fetches data over the network: ['KrisQian-0.0.7/setup.py:40', 'KrisQian-0.0.7/setup.py:50']",
"reads files and dirs: ['KrisQian-0.0.7/setup.py:59', 'KrisQian-0.0.7/setup.py:70']"
]
}
=> Complete report: pypi-KrisQian-0.0.7.json
=> View pre-vetted package report at https://packj.dev/package/PyPi/KrisQian/0.0.7
Packj flagged KrisQian (v0.0.7) as suspicious due to absence of source repo and use of sensitive APIs (network, code generation) during package installation time (in setup.py). We decided to take a deeper look, and found the package malicious. Please find our detailed analysis at https://packj.dev/malware/krisqian.
More examples of malware we found are listed at https://packj.dev/malware Please reach out to us at [email protected] for full list.
To learn more about Packj tool or open-source software supply chain attacks, refer to our
- PyConUS'22 talk and slides.
- BlackHAT Asia'22 Arsenal presentation
- PackagingCon'21 talk and slides
- BlackHat USA'22 Arsenal talk Detecting typo-squatting, backdoored, abandoned, and other "risky" open-source packages using Packj
- Academic dissertation on open-source software security and the paper from our group at Georgia Tech that started this research.
- Open Source Summit, Europe'22 talk Scoring dependencies to detect βweak linksβ in your open-source software supply chain - presentation video on YouTube
- NullCon'22 talk Unearthing Malicious And Other βRiskyβ Open-Source Packages Using Packj
- Add support for other language ecosystems. Rust is a work in progress [ETA: May '23].
- Add functionality to detect several other "risky" code as well as metadata attributes [ETA: Apr '23].
- Self-hosted Packj webserver and several useful integrations (e.g., Gitlab runner) [ETA: Apr'23].
Watch π this repo to stay up to date.
Have a feature or support request? Please visit our GitHub discussion page or join our discord community for discussion and requests.
Packj has been developed by Cybersecurity researchers at Ossillate Inc. and external collaborators to help developers mitigate risks of supply chain attacks when sourcing untrusted third-party open-source software dependencies. We thank our developers and collaborators. Show your appreciation by giving us a β if you like our work.
We welcome code contributions with open arms. See CONTRIBUTING.md guidelines. Found a bug? Please open an issue. Refer to our SECURITY.md guidelines to report a security issue.
What Package Managers (Registries) are supported?
Packj can currently vet NPM, PyPI, and RubyGems packages for "risky" attributes. We are adding support for Rust.
What techniques does Packj employ to detect risky/malicious packages?
Packj uses static code analysis, dynamic tracing, and metadata analysis for comprehensive auditing. Static analysis alone is not sufficient to flag sophisticated malware that can hide itself better using code obfuscation. Dynamic analysis is performed by installing the package under strace
and monitoring it's runtime behavior. Please read more at Audit README.
Does it work on obfuscated calls? For example, a base 64 encrypted string that gets decrypted and then passed to a shell?
This is a very common malicious behavior. Packj detects code obfuscation as well as spawning of shell commands (exec system call). For example, Packj can flag use of getattr()
and eval()
API as they indicate "runtime code generation"; a developer can go and take a deeper look then. See main.py for details.