Talking TUF: Securing
Software Distribution
Justin Cappos, Trishank Kuppusamy, Vladimir Diaz,
Santiago Torres, Sebastien Awwad, Lukas Puehringer
New York University
What do these companies have in common?
What do these companies have in common?
They all had a publicly disclosed software update hack!
Repository compromise impact
● SourceForge mirror distributed malware.
● Attackers impersonate Microsoft Windows Update
to spread Flame malware.
● RubyGems compromised with RCE.
● Opera users automatically installed malware
signed by compromised key.
● Node Packaged Modules compromised.
● Attacks on software updaters have massive impact
○ E.g. South Korea faced 765 million dollars in damages.
Commonly used (bad) techniques
● Why not sign all the software on a community repository?
● This way, we know whether or not attackers have tampered with software after a
repository compromise.
● Couldn’t we already use previous systems --- GPG or TLS --- to do this?
The Problem with TLS
● Good
○ easy to set up
○ has nice lock icon users are trained to trust
● Bad
○ Lots of design / impl issues
○ Compromise repository -> game over
The Problem with GPG
● Good
○ Provides signature of software packages with offline
keys (private keys kept off repository) so that
attackers cannot tamper with packages after a
repository compromise.
● Bad
○ have to manually verify public keys
○ trust for anything usually implies trust for everything
○ Furthermore, only 4% of software projects provide
GPG signatures on PyPI, and 0.07% of users
downloaded GPG signatures between March and
April 2014.
● TUF is a secure software update framework.
● Built on ideas discussed with some folks from Tor.
● Plug-and-play (like TLS), but compromise resilient.
● Goal: support a wide array of different configurations
○ Support, don’t judge!
“Survivable Key Compromise in Software Update
Systems” (CCS 2010).
2010-Present: The Update Framework (TUF)
Design Principles
9
Responsibility
Separation
Multi-signature
Trust
Explicit and
Implicit
Revocation
Minimize
Individual Key
and Role Risk
Design Principles
10
Responsibility
Separation
Delegate roles
to divide
responsibilities
Responsibility Separation
11
Content Timeliness
Design Principles
12
Minimize
Individual Key
and Role Risk
Compromise Risk
=
Probability
x
Impact
Minimize Role & Key Risk
13
Root
High-impact role? => Highly-secure keys
Timeliness
Online keys? => Low-impact role
Design Principles
14
Multi-signature
Trust
(t, n)
signature threshold
required for trust
Multi-signature Trust
15
A
B
A
No risk to clients.
Signature threshold:
Two signatures
Design Principles
16
Explicit and
Implicit
Revocation
Explicit and Implicit Revocation
17
A
C
B
Signature threshold:
Two signatures
A
B
B
A
18
Design
19
Root
Targets
(projects) TimestampSnapshot
Malware attack
django-1.7.1.tar.gz
bcrypt-1.1.1.tar.gz
flask-0.10.tar.gz
django
bcrypt
flask
django-1.8.tar.gz
repository (compromised)
user
malware!
Versions of metadata
django-1.7.1.tar.gzdjango
metadata
version
developers packages
● packages
○ django-1.7.1.tar.gz
■ hash: X
● version: 1
Just as there as different
versions of packages...
Versions of metadata
django-1.7.1.tar.gz
django django-1.8.tar.gz
metadata
version
developers packages
● packages
○ django-1.8.tar.gz
■ hash: Y
○ django-1.7.1.tar.gz
■ hash: X
● version: 2
...there are different
versions of metadata
corresponding to
different versions of
packages.
The version number
of a metadata file (e.g.
2) does not
correspond with the
version number of
packages (e.g. 1.7.1).
Replay attack
version
package
django bcrypt flask
4
5
2
version
package
django bcrypt flask
3
2
1
replay!
old & vulnerable!
TUF: eager verification
django-1.7.1.tar.gz
bcrypt-1.1.1.tar.gz
flask-0.10.tar.gz
django
bcrypt
flask
django-1.8.tar.gz
repository
user
developer
metadata
snapshot
administrator
metadata
hash
hash
hash
version
version
version
1
2
3
5
4
User downloads
all package
metadata to
verify snapshot
metadata.
Why? To prevent
replay attacks,
and not blindly
trust
administrators.
TUF: snapshot
● Adds a “snapshot” of all metadata/packages.
version
package
django bcrypt flask
4
5
2
packages not installed,
but metadata downloaded version
package
django bcrypt flask
4
2
1
packages installed,
but with obsolete metadata
replay!
Secure lazy verification
django-1.7.1.tar.gz
bcrypt-1.1.1.tar.gz
flask-0.10.tar.gz
django
bcrypt
flask
django-1.8.tar.gz
repository
user
developer
metadata
snapshot
administrator
metadata
version
version
version
version
version
version
1
2
3
User downloads
only snapshot +
desired package
metadata!
Trust
administrators to
specify accurate
snapshot
metadata.
Version checking
● Compact “snapshot” of all metadata/packages.
version
package
django bcrypt flask
4
5
2
packages not installed,
but version downloaded version
package
django bcrypt flask
4
2
1
packages installed,
but with obsolete metadata
replay!
Is this as secure as hash checking?
● So what security attacks have we given up?
○ Not malware attacks, because package metadata
still signed with offline developer keys.
○ Not replay attacks, because snapshot metadata
cannot specify older version numbers.
Fast-forward attack
version
package
django bcrypt flask
4
5000
2000
packages not installed,
but version downloaded version
package
django bcrypt flask
4
5
2
packages not installed,
due to version mismatch
denied!
Only a mild,
denial-of-service
attack.
Okay, but is it as secure as hash
checking?
Yes!
● FF DoS (~= dropping requests)
○ Address by resetting version numbers after key
revocation.
Example setup for TUF
1. Responsibility separation (roles)
2. Multitrust signatures (a.k.a. two-man rule).
a. some roles like root may need multiple signatures from keys
3. Explicit and implicit revocation of keys.
a. individual roles / keys timeout
4. Minimizing risk (with offline keys).
5. Further selective delegation from targets role.
a. Gives trust without sharing keys, etc.
ε
timestamp
metadata packages
online
keys
offline
keys
signs metadata for
target
package
signs root keys for
delegates packages to
root
snapshot targets
A1
BC
A.pkg
C.gz
signs for packages
A.*B.*,C.*
*.pkg
A2
B.tar
Multi-trust signatures
● Can require multiple signatures for a role
○ Some keys can be lost / compromised and things work
>>> repository = create_new_repository("repository/")
>>> public_root_key = import_rsa_publickey_from_file("keystore/root_key.pub")
>>> repository.root.add_verification_key(public_root_key)
>>> public_root_key2 = import_rsa_publickey_from_file("keystore/root_key2.pub")
>>> repository.root.add_verification_key(public_root_key2)
# Threshold of each role defaults to 1.
>>> repository.root.threshold
1
# Set threshold then need to write / sign the new root file.
>>> repository.root.threshold = 2
>>> repository.root.load_signing_key(private_root_key)
>>> repository.root.load_signing_key(private_root_key2)
>>> repository.writeall()
Target (Project) Delegation in PyPI (PEP 480)
● Lots of good suggestions for changes to TUF
● Formal TUF Augmentation Proposal (TAP) process
○ Discuss ideas, when ‘close’ send TAP
○ We review closely
○ Test implementation
○ Approve
○ (Read TAPs 1 and 2 for details)
https://github.com/theupdateframework/taps/blob/master/tap1.md
Standardization process (TAPs)
● TAP 3 -- multi-role signatures (Evan / Jake)
○ Alice AND Bob must both sign package A
○ Lets one have ‘unequal’ quorums
● TAP 4 -- pinning repository keys (Evan / Jake)
○ The user can control the root of trust for parts of the
namespace
■ Root role compromise !-> game over!
● TAP 5 -- specify URLs in root files
○ Makes it easy to change the repo location
● TAP 6 -- version numbers in root metadata (David)
● TAP ? -- hash chaining of timestamp metadata (???)
○ Coming soon?
https://github.com/theupdateframework/taps/blob/master/tap1.md
Standardization process (TAPs cont...)
Integrations of TUF (some on-going)
Related effort: Uptane (securing
automotive software updates)
Uptane: Securely updating automobiles
Work closely with vendors, OEMs, etc.
● Security reps from 79% of US cars
● Many top suppliers / vendors
Account for deployment concerns
● Solutions are only useful if deployed
● Accommodate existing infrastructure,
business relationships, etc.
Standardize and harden
● Working toward SAE certification
● Professional security audit
● Free / open source, detailed tests /
Uptane: Securely updating automobiles
Current design
Latest downloaded
metadata
Latest downloaded
encrypted image
Boot-
loader
Previous
metadata
ECU
keys
Uptane Timeline
40
● Current tasks:
○ High level spec (complete!)
○ Multi-group security analysis (complete!)
○ Detailed impl specification (RFC-style) (?complete??)
○ Reference implementation (in progress)
○ Compliance test cases (in progress)
○ Deployment recommendations document (in progress)
● Upcoming:
○ Technology demonstration (Oct 18)
○ Public security review
○ SAE Standardization
Future work: healthcare, infrastructure too
Healthcare systems:
● Often antiquated OSes / systems
● Only certified in a specific configurations
● Increasingly targeted
Infrastructure:
● Often antiquated OSes / systems
● Reliability is the focus, not security
○ Remote access needed
Security issues can have catastrophic impact!
Related effort: Toto (securing the
software supply chain)
43
Toto
Toto: Overview
Project owner Functionaries End User
What needs to be done Perform steps, provide
evidence
Verify
Layout
Link
Link
Link
Link
Link
Final
Product
Toto: Overview
Project owner
Defines the steps that are required in this project’s software
supply chain
Layout
● Only Alice and Bob can commit to
this VCS
● The build will be made using the
company’s Gradle buildserver
● The project will be added to a
docker recipe by Carl
● ...
Toto: OverviewFunctionaries
Perform steps and provide evidence as link metadata
Link
Link
Link
● Alice: I committed to the VCS
● Gradle buildserver: I compiled
alice’s commit
● Carl: I pulled and made a docker
image of all of this
Toto: Overview
End user
Verifies the metadata
Link
Link
Link
Link
Link
Final
Product
Layout
Timeline
49
● Currently:
○ High level spec (release coming ~1 week)
○ Reference implementation (“complete” ~1-2 weeks)
● Upcoming:
○ Internal use (~2-3 weeks)
○ Compliance test cases (~3 weeks)
○ External beta testing (~1-2 mo)
○ Broad public release (???)
Wrapping up
Conclusion
51
● Securing software distribution, etc. is hard
● Notary provides strong guarantees for Docker containers
● Use TAPs to get changes into TUF (let’s discuss first)
● Let’s work together!
○ https://github.com/theupdateframework/
○ https://github.com/uptane
○ https://github.com/toto-framework
Thanks!
Questions?
https://theupdateframework.com
https://isis.poly.edu/~jcappos/
jcappos@nyu.edu
My background... (2003-2008)
● Built the first package manager designed specifically for OSVMs (Stork)
○ Deployed on the research infrastructure “PlanetLab”
■ Practical experience: thousands of VM instances over 8 years of use
○ Packages are cached in a special VM and shared
■ Disk, memory, and bandwidth savings
■ Additional security risks [USENIX ATC 2005], [LISA 2007]
2008: Attacks on Linux package managers
● By changing unsigned metadata, we can compromise users.
● No protection against:
○ Arbitrary package attacks
○ Extraneous dependencies
○ Replay attacks
○ Mix-and-match attacks
“A Look in the Mirror: Attacks on Package Managers”
(CCS 2008).
Fixing Linux package managers
● Disclosed these security attacks via CERT (VU#230187).
● Major vendors have adopted our security architecture.
2009: Mission accomplished!
...or is it???
2009: Tor
● Tor: “We heard about your work. Can you help us fix our software
updater?”
● Security is simple, right?
● How hard can this be anyway?
Thandy (Tor)
● The Thandy software updater for
Tor
○ A quorum of keys for root of trust.
○ Signing by different
compartmentalized key types.
○ Use online keys only to prevent freeze
attacks and
bound trust window.
Thandy (Tor)
● The Thandy software updater for
Tor
○ A quorum of keys for root of trust.
○ Signing by different
compartmentalized key types.
○ Use online keys only to prevent freeze
attacks and
bound trust window.
○ ...still not enough.
● Still found 8 security problems.
● Building your own secure software
updater is not trivial.

Talking TUF: Securing Software Distribution

  • 1.
    Talking TUF: Securing SoftwareDistribution Justin Cappos, Trishank Kuppusamy, Vladimir Diaz, Santiago Torres, Sebastien Awwad, Lukas Puehringer New York University
  • 2.
    What do thesecompanies have in common?
  • 3.
    What do thesecompanies have in common? They all had a publicly disclosed software update hack!
  • 4.
    Repository compromise impact ●SourceForge mirror distributed malware. ● Attackers impersonate Microsoft Windows Update to spread Flame malware. ● RubyGems compromised with RCE. ● Opera users automatically installed malware signed by compromised key. ● Node Packaged Modules compromised. ● Attacks on software updaters have massive impact ○ E.g. South Korea faced 765 million dollars in damages.
  • 5.
    Commonly used (bad)techniques ● Why not sign all the software on a community repository? ● This way, we know whether or not attackers have tampered with software after a repository compromise. ● Couldn’t we already use previous systems --- GPG or TLS --- to do this?
  • 6.
    The Problem withTLS ● Good ○ easy to set up ○ has nice lock icon users are trained to trust ● Bad ○ Lots of design / impl issues ○ Compromise repository -> game over
  • 7.
    The Problem withGPG ● Good ○ Provides signature of software packages with offline keys (private keys kept off repository) so that attackers cannot tamper with packages after a repository compromise. ● Bad ○ have to manually verify public keys ○ trust for anything usually implies trust for everything ○ Furthermore, only 4% of software projects provide GPG signatures on PyPI, and 0.07% of users downloaded GPG signatures between March and April 2014.
  • 8.
    ● TUF isa secure software update framework. ● Built on ideas discussed with some folks from Tor. ● Plug-and-play (like TLS), but compromise resilient. ● Goal: support a wide array of different configurations ○ Support, don’t judge! “Survivable Key Compromise in Software Update Systems” (CCS 2010). 2010-Present: The Update Framework (TUF)
  • 9.
  • 10.
  • 11.
  • 12.
    Design Principles 12 Minimize Individual Key andRole Risk Compromise Risk = Probability x Impact
  • 13.
    Minimize Role &Key Risk 13 Root High-impact role? => Highly-secure keys Timeliness Online keys? => Low-impact role
  • 14.
  • 15.
    Multi-signature Trust 15 A B A No riskto clients. Signature threshold: Two signatures
  • 16.
  • 17.
    Explicit and ImplicitRevocation 17 A C B Signature threshold: Two signatures A B B A
  • 18.
  • 19.
  • 20.
  • 21.
    Versions of metadata django-1.7.1.tar.gzdjango metadata version developerspackages ● packages ○ django-1.7.1.tar.gz ■ hash: X ● version: 1 Just as there as different versions of packages...
  • 22.
    Versions of metadata django-1.7.1.tar.gz djangodjango-1.8.tar.gz metadata version developers packages ● packages ○ django-1.8.tar.gz ■ hash: Y ○ django-1.7.1.tar.gz ■ hash: X ● version: 2 ...there are different versions of metadata corresponding to different versions of packages. The version number of a metadata file (e.g. 2) does not correspond with the version number of packages (e.g. 1.7.1).
  • 23.
    Replay attack version package django bcryptflask 4 5 2 version package django bcrypt flask 3 2 1 replay! old & vulnerable!
  • 24.
  • 25.
    TUF: snapshot ● Addsa “snapshot” of all metadata/packages. version package django bcrypt flask 4 5 2 packages not installed, but metadata downloaded version package django bcrypt flask 4 2 1 packages installed, but with obsolete metadata replay!
  • 26.
  • 27.
    Version checking ● Compact“snapshot” of all metadata/packages. version package django bcrypt flask 4 5 2 packages not installed, but version downloaded version package django bcrypt flask 4 2 1 packages installed, but with obsolete metadata replay!
  • 28.
    Is this assecure as hash checking? ● So what security attacks have we given up? ○ Not malware attacks, because package metadata still signed with offline developer keys. ○ Not replay attacks, because snapshot metadata cannot specify older version numbers.
  • 29.
    Fast-forward attack version package django bcryptflask 4 5000 2000 packages not installed, but version downloaded version package django bcrypt flask 4 5 2 packages not installed, due to version mismatch denied! Only a mild, denial-of-service attack.
  • 30.
    Okay, but isit as secure as hash checking? Yes! ● FF DoS (~= dropping requests) ○ Address by resetting version numbers after key revocation.
  • 31.
    Example setup forTUF 1. Responsibility separation (roles) 2. Multitrust signatures (a.k.a. two-man rule). a. some roles like root may need multiple signatures from keys 3. Explicit and implicit revocation of keys. a. individual roles / keys timeout 4. Minimizing risk (with offline keys). 5. Further selective delegation from targets role. a. Gives trust without sharing keys, etc. ε timestamp metadata packages online keys offline keys signs metadata for target package signs root keys for delegates packages to root snapshot targets A1 BC A.pkg C.gz signs for packages A.*B.*,C.* *.pkg A2 B.tar
  • 32.
    Multi-trust signatures ● Canrequire multiple signatures for a role ○ Some keys can be lost / compromised and things work >>> repository = create_new_repository("repository/") >>> public_root_key = import_rsa_publickey_from_file("keystore/root_key.pub") >>> repository.root.add_verification_key(public_root_key) >>> public_root_key2 = import_rsa_publickey_from_file("keystore/root_key2.pub") >>> repository.root.add_verification_key(public_root_key2) # Threshold of each role defaults to 1. >>> repository.root.threshold 1 # Set threshold then need to write / sign the new root file. >>> repository.root.threshold = 2 >>> repository.root.load_signing_key(private_root_key) >>> repository.root.load_signing_key(private_root_key2) >>> repository.writeall()
  • 33.
    Target (Project) Delegationin PyPI (PEP 480)
  • 34.
    ● Lots ofgood suggestions for changes to TUF ● Formal TUF Augmentation Proposal (TAP) process ○ Discuss ideas, when ‘close’ send TAP ○ We review closely ○ Test implementation ○ Approve ○ (Read TAPs 1 and 2 for details) https://github.com/theupdateframework/taps/blob/master/tap1.md Standardization process (TAPs)
  • 35.
    ● TAP 3-- multi-role signatures (Evan / Jake) ○ Alice AND Bob must both sign package A ○ Lets one have ‘unequal’ quorums ● TAP 4 -- pinning repository keys (Evan / Jake) ○ The user can control the root of trust for parts of the namespace ■ Root role compromise !-> game over! ● TAP 5 -- specify URLs in root files ○ Makes it easy to change the repo location ● TAP 6 -- version numbers in root metadata (David) ● TAP ? -- hash chaining of timestamp metadata (???) ○ Coming soon? https://github.com/theupdateframework/taps/blob/master/tap1.md Standardization process (TAPs cont...)
  • 36.
    Integrations of TUF(some on-going)
  • 37.
    Related effort: Uptane(securing automotive software updates)
  • 38.
    Uptane: Securely updatingautomobiles Work closely with vendors, OEMs, etc. ● Security reps from 79% of US cars ● Many top suppliers / vendors Account for deployment concerns ● Solutions are only useful if deployed ● Accommodate existing infrastructure, business relationships, etc. Standardize and harden ● Working toward SAE certification ● Professional security audit ● Free / open source, detailed tests /
  • 39.
    Uptane: Securely updatingautomobiles Current design Latest downloaded metadata Latest downloaded encrypted image Boot- loader Previous metadata ECU keys
  • 40.
    Uptane Timeline 40 ● Currenttasks: ○ High level spec (complete!) ○ Multi-group security analysis (complete!) ○ Detailed impl specification (RFC-style) (?complete??) ○ Reference implementation (in progress) ○ Compliance test cases (in progress) ○ Deployment recommendations document (in progress) ● Upcoming: ○ Technology demonstration (Oct 18) ○ Public security review ○ SAE Standardization
  • 41.
    Future work: healthcare,infrastructure too Healthcare systems: ● Often antiquated OSes / systems ● Only certified in a specific configurations ● Increasingly targeted Infrastructure: ● Often antiquated OSes / systems ● Reliability is the focus, not security ○ Remote access needed Security issues can have catastrophic impact!
  • 42.
    Related effort: Toto(securing the software supply chain)
  • 43.
  • 44.
  • 45.
    Toto: Overview Project ownerFunctionaries End User What needs to be done Perform steps, provide evidence Verify Layout Link Link Link Link Link Final Product
  • 46.
    Toto: Overview Project owner Definesthe steps that are required in this project’s software supply chain Layout ● Only Alice and Bob can commit to this VCS ● The build will be made using the company’s Gradle buildserver ● The project will be added to a docker recipe by Carl ● ...
  • 47.
    Toto: OverviewFunctionaries Perform stepsand provide evidence as link metadata Link Link Link ● Alice: I committed to the VCS ● Gradle buildserver: I compiled alice’s commit ● Carl: I pulled and made a docker image of all of this
  • 48.
    Toto: Overview End user Verifiesthe metadata Link Link Link Link Link Final Product Layout
  • 49.
    Timeline 49 ● Currently: ○ Highlevel spec (release coming ~1 week) ○ Reference implementation (“complete” ~1-2 weeks) ● Upcoming: ○ Internal use (~2-3 weeks) ○ Compliance test cases (~3 weeks) ○ External beta testing (~1-2 mo) ○ Broad public release (???)
  • 50.
  • 51.
    Conclusion 51 ● Securing softwaredistribution, etc. is hard ● Notary provides strong guarantees for Docker containers ● Use TAPs to get changes into TUF (let’s discuss first) ● Let’s work together! ○ https://github.com/theupdateframework/ ○ https://github.com/uptane ○ https://github.com/toto-framework
  • 52.
  • 53.
    My background... (2003-2008) ●Built the first package manager designed specifically for OSVMs (Stork) ○ Deployed on the research infrastructure “PlanetLab” ■ Practical experience: thousands of VM instances over 8 years of use ○ Packages are cached in a special VM and shared ■ Disk, memory, and bandwidth savings ■ Additional security risks [USENIX ATC 2005], [LISA 2007]
  • 54.
    2008: Attacks onLinux package managers ● By changing unsigned metadata, we can compromise users. ● No protection against: ○ Arbitrary package attacks ○ Extraneous dependencies ○ Replay attacks ○ Mix-and-match attacks “A Look in the Mirror: Attacks on Package Managers” (CCS 2008).
  • 55.
    Fixing Linux packagemanagers ● Disclosed these security attacks via CERT (VU#230187). ● Major vendors have adopted our security architecture.
  • 56.
  • 57.
    2009: Tor ● Tor:“We heard about your work. Can you help us fix our software updater?” ● Security is simple, right? ● How hard can this be anyway?
  • 58.
    Thandy (Tor) ● TheThandy software updater for Tor ○ A quorum of keys for root of trust. ○ Signing by different compartmentalized key types. ○ Use online keys only to prevent freeze attacks and bound trust window.
  • 59.
    Thandy (Tor) ● TheThandy software updater for Tor ○ A quorum of keys for root of trust. ○ Signing by different compartmentalized key types. ○ Use online keys only to prevent freeze attacks and bound trust window. ○ ...still not enough. ● Still found 8 security problems. ● Building your own secure software updater is not trivial.