Tor Stinks

Open PDF
Creating agency: NSA National Security Agency (US)
Document date: 2012-06
Pages: 23
Genre: slides
Banner:
Classification: TOP SECRET COMINT
Dissemination: REL FVEY

A June 2012 joint NSA-GCHQ presentation titled "Tor Stinks" detailing the two agencies' attempts to de-anonymize Tor users through analytics, exploitation, and network disruption — candidly admitting they will never be able to identify all Tor users, but outlining a range of techniques including cookie leakage tracking, QUANTUM attacks, circuit reconstruction, timing analysis, and deliberate network degradation.

Long Description:
This top secret NSA presentation, produced by the CT SIGDEV unit in June 2012 and shared with Five Eyes partners, documents the state of joint NSA-GCHQ efforts to attack the Tor anonymity network as of that date. The document opens with a frank admission: the agencies will never be able to de-anonymize all Tor users all the time, and have had no success de-anonymizing users on demand in response to specific TOPI requests. The presentation emerged from REMATION II, a joint NSA-GCHQ counter-Tor workshop split between NSA's MHS facility (analytics focus) and GCHQ (exploitation focus), and organizes its findings across four categories: analytics, technical research, exploitation, and Tor node operations. The analytics section covers six approaches. Circuit reconstruction attempts to identify all three nodes in a Tor circuit simultaneously — currently limited by the agencies' access to very few nodes, making success negligible. The "Goes Inta Goes Outta" low-latency correlation technique, for which GCHQ has a working version called QUICKANT, attempts to correlate traffic patterns entering and exiting the Tor network to identify users. Cookie leakage exploits the fact that some cookies survive Tor use, with proposals to seed cookies to specific targets and investigate Evercookie persistence — including observed cases of DoubleClick IDs appearing on both Tor and non-Tor IPs. The EPICFAIL/GCHQ QFD and NSA's proposed GREAT EXPECTATIONS variant identify users through hard selectors when they are not using Tor. Node lifespan tracking via RONIN and DNS behavior analysis round out the analytics work. The exploitation section is the most operationally significant. QUANTUMCOOKIE forces Tor clients to reveal stored browser cookies — including Hotmail MUIDs and Yahoo cookies — by intercepting an exit node request and redirecting the client to major web services that then serve identifying cookies in the clear. The presentation also discusses using FOXACID (referred to as "FA") and SHORTSHEET CNE techniques against Tor Browser Bundle users, though noting that Torbutton and TBB currently prevent CNE success against protected configurations. Shaping operations aim to route users toward friendly exit nodes, stain their traffic, or instruct compromised computers to connect outside Tor and reveal their true IP. Web server enabling would modify compromised servers to execute timing attacks. The node flooding concept proposes deploying slow nodes advertised as high-bandwidth to degrade overall network stability. The document closes on a pragmatic note — acknowledging that scaring targets away from Tor could be counterproductive, since a critical mass of targets use it, and that the goal is increasing success rates for individual users rather than total de-anonymization.

TOP SECRET//COMINT// REL FVEY

Tor Stinks (U)

[REDACTED]
CT SIGDEV
[REDACTED]
JUN 2012

Derived From: NSA/CSSM 1-52
Dated: 20070108
Declassify On: 20370101

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Tor Stinks... (U)

  • We will never be able to de-anonymize all Tor users all the time.
  • With manual analysis we can de-anonymize a very small fraction of Tor users, however, no success de-anonymizing a user in response to a TOPI request/on demand.

[REDACTED]

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

REMATION II (U)

  • Joint NSA GCHQ counter-Tor workshop
  • Week one at MHS focus on analytics
  • Week two at GCHQ focus on exploitation

https://wiki.gchq/index.php?title=REMATION

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Laundry List (U)

  • Analytics to de-anonymize users
    • Circuit reconstruction (21)
    • Goes inta goes outta/low latency (2)
    • Cookie leakage
    • Dumb users (EPICFAIL)
    • Node Lifespan (17)
    • DNS
  • Technical Analysis/Research
    • Hidden services (4, 5, 6, 7)
    • Timing pattern (3)
    • Torservers.net/Amazon AWS
  • Exploitation
    • QUANTUM attacks (1, 20, 22)
    • Existing options (8 + 11)
    • Shaping (9 + 16)
    • Web server enabling (10)
    • Nodes (14)
    • Degrade user experience (13 + 18)
  • Nodes
    • Baseline our nodes (21)
    • Tor node flooding

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: Circuit Reconstruction (S//SI)

Terrorist with Tor client installed
Tor entry node
Tor relay node
Tor exit node
Internet site
ANONYMIZER CLOUD

  • Current: access to very few nodes. Success rate negligible because all three Tor nodes in the circuit have to be in the set of nodes we have access to.
    • Difficult to combine meaningfully with passive SIGINT.
  • Goal: expand number of nodes we have access to
    • GCHQ runs Tor nodes under NEWTONS CRADLE (how many?)
    • Other partners?
    • Partial reconstruction (first hops or last hops)?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics:

Goes Inta Goes Outta/Low Latency (S//SI)

Find possible alternative accounts for a target: look for connections to Tor, from the target's suspected country, near time of target's activity.

  • Current: GCHQ has working version (QUICKANT). R has alpha tested NSA's version. NSA's version produced no obvious candidate selectors.
  • Goal: Figure out if QUICKANT works, compare methodologies. Gathering data for additional tests of NSA's version (consistent, random and heavy user)

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: Cookie Leakage (TS//SI)

Use cookies to identify Tor users when they are not using Tor

  • Current: preliminary analysis shows that some cookies “survive” Tor use. Depends on how target is using Tor (Torbutton/Tor Browser Bundle clears out cookies).
  • Goal: test with cookies associated with CT targets
    • Idea: what if we seeded cookies to a target?
    • Investigate Evercookie persistence

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: Cookie Leakage (TS//SI)

  • DoubleclickID seen on Tor and non-Tor IPs

[REDACTED]

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: Dumb Users (EPICFAIL) (S//SI)

GCHQ QFD that looks for Tor users when they are not using Tor.

  • Current: GCHQ has working QFD based on hard selector (email, web forum, etc) but does not include cookies.
  • Goal: NSA investigating own version (GREAT EXPECTATIONS) that would include cookies.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: Node Lifespan (S//SI)

How do I know WHEN a particular IP was a Tor node as opposed to IF it was a Tor node?

  • Current: detection done once an hour by NTOC. RONIN stores "last seen" and nodes age off slowly with no accurate lifespan.
  • Goal: Working with RONIN to add more details on node lifespan.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Analytics: DNS (TS//SI)

How does Tor handle DNS requests? Are DNS requests going through Tor? Does this depend on how the target is using Tor?

  • Current: Still investigating.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Technical Analysis: Hidden Services (TS//SI)

What do we know about Hidden Services?

  • Current: No effort by NSA, some DSD and GCHQ work on ONIONBREATH.
  • Goal:
    • Harvest and enumerate .onion URLs
    • Identify similar HS based on referrer fields
    • Distinguish HS from normal Tor clients

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Technical Analysis: Timing Pattern (TS//SI)

Send packets back to the client that are detectable by passive accesses to find client IPs for Tor users.

  • Current: GCHQ has research paper and demonstrated capability in the lab.
  • Goal: Can we expand to other owned nodes?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Technical Analysis: torservers.net (TS//SI)

Investigate the Amazon AWS cloud instances of Tor servers. How are IPs allocated and reassigned once bandwidth limit is reached? Impact on RONIN's ability to detect nodes?

  • Current: GCHQ set up Tor nodes on the AWS cloud during REMATION II.

[REDACTED]

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: QUANTUM (TS//SI)

  • QUANTUM to degrade/deny/disrupt Tor access?
  • QUANTUMCOOKIE – forces clients to divulge stored cookies.
STAGE 1
Tor Client (running the Tor browser bundle)
A client visits a website via TOR

Encrypted GET Request to www.terrorist.com/bomb.php

TOR Network

In the clear GET Request to www.terrorist.com/bomb.php
www.terrorist.com

We detect the GET request and respond with a redirect to Hotmail and Yahoo!
QUANTUM

STAGE 2
Tor Client (running the Tor browser bundle)
The client now sends two GET requests. One to Hotmail and one to Yahoo! Corresponding Cookies will be sent if stored on the client's browser

Encrypted GET Request to Hotmail.com with MUID
Encrypted GET Request to Yahoo.com with Y-Cookie

TOR Network

In the clear GET Request to Hotmail.com with MUID
Hotmail

In the clear GET Request to Yahoo.com with Y-Cookie
Yahoo!

Figure 4. A diagram of how the QUANTUM Survey / Cookie technique works

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: Existing Options (TS//SI)

Test current CNE techniques (FA and SHORTSHEET) against Torbutton and TBB users.

  • Current: Torbutton and TBB prevent CNE success. Possible success against "vanilla" Tor/Vidalia.
  • Goal: modifications to initial CNE surveys? Ignore user-agents from Torbutton or TBB? Improve browser fingerprinting? Using javascript instead of Flash?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: Shaping (TS//SI)

  • Given CNE access to a target computer can we shape their traffic to "friendly" exit nodes?
  • Route users to a separate “private” Tor network?
  • Stain their traffic or user agent?
  • Instruct target computer to use a service that connects outside Tor and reveal true IP?
  • Current: Can stain user agent working on shaping.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: Web Server Enabling (TS//SI)

Given CNE access to web server modify the server to enable a “timing/counting” attack similar to timing pattern idea.

  • Current: GCHQ has a research paper and demonstrated the technique in the lab.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: Nodes (TS//SI)

Can we exploit nodes?
Probably not. Legal and technical challenges.

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Exploitation: Degrade Tor experience (TS//SI)

Given CNE access to a web server make it painful for Tor users?

Given CNE access to a network can we deny/degrade/disrupt Tor users' ?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Nodes: Baseline Our Nodes (TS//SI)

How many nodes do we have cooperative or direct access to? Can we deploy similar code to these nodes to aid with circuit reconstruction?

Can we do packet timing attacks using nodes?
Can we use the nodes to shape traffic flow?
Can we use the nodes to deny/degrade/disrupt comms to certain sites?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Nodes: Tor Node Flooding (TS//SI)

Could we set up a lot of really slow Tor nodes (advertised as high bandwidth) to degrade the overall stability of the network?

TOP SECRET//COMINT// REL FVEY


TOP SECRET//COMINT// REL FVEY

Tor Stinks... But it Could be Worse (S//SI)

  • Critical mass of targets use Tor. Scaring them away from Tor might be counterproductive.
  • We can increase our success rate and provide more client IPs for individual Tor users.
  • Will never get 100% but we don't need to provide true IPs for every target every time they use Tor.

TOP SECRET//COMINT// REL FVEY