why hackers don’t get caught

Network: hiding the address

Every packet carries a return address; hiding it means handing your traffic to someone else, and the question is always who that someone is and what they can see.

What it hides
The real IP address and who talks to whom
How it gives way
One connection without the tunnel, or timing correlation
How much to read

A packet cannot be delivered without a source address, so every server you touch learns where to send the reply. Someone who wants to act without being found has to make that address belong to somebody else: a VPN company, a stranger’s router, or three volunteers on the Tor network. Each option changes who holds the missing piece, and none of them makes the piece disappear.

What an address gives away

On its own an IP address is a number. Joined with a timestamp, it becomes a question an investigator can send to the internet service provider that was assigned that address: which subscriber held it at that moment. Behind a carrier-grade NAT the provider needs the source port as well, but the principle is the same. The provider’s assignment records turn a number in a server log into a name and a billing address.

That is why the address is the thing people try to hide, and also why courts treat it as information the user hands out. In the Silk Road 2.0 related case against Brian Farrell, a US federal judge wrote in 2016 that Tor users “clearly lack a reasonable expectation of privacy in their IP addresses”, because they disclose them to the network in order to use it.1

The interactive figure on the home page lets you pick an observer on the path and see which fields of the record it can read. This chapter looks at the tools that move the address, one at a time, and at what each tool asks you to trust.

One hop: a VPN moves the trust, it does not remove it

A VPN or a proxy is a single relay. The destination sees the relay’s address instead of yours. Your ISP sees an encrypted tunnel to the relay. The relay operator sees both ends: your real address on one side, and every destination on the other. Anonymity from the destination is bought by giving one company the complete picture. Tor’s own support pages make the same point when they say they do not recommend combining a VPN with Tor unless the user knows how to configure both without compromising privacy.2

Whether that company keeps the picture is a matter of policy, not mechanism, and policies have failed in court records. In 2017 the FBI investigated a cyberstalking campaign run through PureVPN, which advertised that it kept no logs that could identify users. The affidavit states that PureVPN “was able to determine that their service was accessed by the same customer from two originating IP addresses”: the suspect’s home connection and his employer’s network.34 The provider had no record of what he did, but it had connection records, and that was enough. In a 2016 child-exploitation investigation that became public in 2018, Highwinds, then the owner of IPVanish, answered a second summons with the subscriber’s name, email address and connection times, which led investigators to the suspect’s home ISP.5

Borrowed addresses: residential proxies and router botnets

The destination does not only ask where traffic comes from. It also asks what kind of address it is. Datacenter and VPN ranges are easy to flag; a home broadband address looks like a customer. So a market grew for residential addresses, and much of the supply comes from devices whose owners never agreed to relay anyone’s traffic.

The 911 S5 botnet is the largest documented example. According to the US Justice Department, between 2014 and July 2022 it infected millions of Windows computers, associated with more than 19 million unique IP addresses, largely through free VPN applications its administrator distributed (MaskVPN and DewVPN) and through bundles with pirated software. Access to those addresses was sold to customers who used them, among other things, to file about 560,000 fraudulent unemployment claims. The administrator allegedly made about $99 million.6 People who installed a free VPN to hide their own address became the exit point for other people’s fraud.

State operators use the same trick with routers. In January 2024 the Justice Department announced a court-authorized operation against the KV-botnet, which it described as hundreds of US small-office and home routers, mostly end-of-life Cisco and NETGEAR models, that Volt Typhoon used to “conceal the PRC origin” of intrusions into critical infrastructure.7 A joint advisory in September 2024 described Raptor Train, a botnet of more than 260,000 devices as of June 2024, controlled by the Beijing company Integrity Technology Group.8 Two weeks after the KV-botnet announcement, the FBI cleaned a botnet of Ubiquiti routers that criminals had built with default passwords and that Russian military intelligence had then used for its own operations.9 One set of hijacked routers served two very different customers.

Mandiant calls the organised version of this an ORB network (operational relay box): a mesh of leased virtual servers and compromised, often end-of-life routers and IoT devices, administered by contractors and rented to several espionage groups at once. The traffic leaves through exit nodes close to the target, so it “blends in” for the analysts reviewing it, and an address can stay in the network for as few as 31 days.10

Operator serverRelay VPSTraversal nodesExit nodeVictimDefender sees198.51.100.143a home router nearbyNode lifetimes (illustrative)on the blocklistnot on the blocklistblocklist built, day 40traffic on day 75198.51.100.14203.0.113.61198.51.100.90203.0.113.7198.51.100.143203.0.113.201198.51.100.36day 0day 30day 60day 90day 120
Why IP blocklists age quickly against an ORB network. Top: the defender sees only the exit node. Bottom: a blocklist built from addresses seen by day 40 no longer matches any node carrying traffic on day 75. Hatched spans are on the blocklist; the yellow row is the exit node shown above. Node lifetimes are illustrative; Mandiant reports addresses lasting as few as 31 days.

For a defender, the consequence is that the source address answers almost nothing. It is often a router in the defender’s own country, belonging to someone who has never heard of the operation. Attribution has to move to behaviour and to infrastructure patterns observed over time, which is the subject of the next chapter.

Tor: three relays, three layers

Tor removes the single party who sees both ends. The client picks three relays, a guard, a middle and an exit, and builds the circuit one hop at a time, negotiating a separate set of keys with each. In the words of the Tor Project, “each relay along the way knows only which relay gave it data and which relay it is giving data to”.11 Tor moves data in small fixed-size packets it calls cells. Every cell the client sends is wrapped three times, once per relay key, like a letter sealed in three envelopes. Each relay opens the one envelope addressed to it, reads where to send the rest, and passes on what it cannot open.

1. You198.51.100.23wraps the request three timesguard layermiddle layerexit layerrequest to forum.exampleWhat this point can readSender address198.51.100.23 (you)Passes it to192.0.2.14Destinationforum.exampleContentyour request2. Guard192.0.2.14opens the guard layermiddle layerexit layerWhat this point can readSender address198.51.100.23 (you)Passes it to192.0.2.77 (middle)DestinationContent3. Middle192.0.2.77opens the middle layerexit layerWhat this point can readSender address192.0.2.14 (guard)Passes it to192.0.2.201 (exit)DestinationContent4. Exit192.0.2.201opens the last layer; TLS remainsrequest to forum.exampleWhat this point can readSender address192.0.2.77 (middle)Passes it toforum.exampleDestinationforum.exampleContent
One request at four points on a Tor circuit. Each relay opens only the envelope sealed to its own key, which tells it the next hop and nothing more. Only the guard sees your address (magenta); only the exit sees the destination. Between the exit and forum.example the content is still protected by TLS.

Several details matter for how this fails or holds:

  • Guards are sticky. If a client picked a fresh entry relay for each circuit, a hostile relay would eventually be its entry and see its address. Tor instead keeps a small set of guards for a long period, so a client either has a bad guard or does not.12
  • Circuits are reused for about ten minutes and then replaced, so connections close in time share an exit.11
  • The relay list is a signed consensus. A small fixed set of directory authorities vote every hour on which relays exist and which flags they hold, and clients accept a consensus signed by more than half of them.13 In early November 2026 the network had about 9,500 relays and 2,500 bridges.14 User counts are estimates derived from directory requests, a few million a day, and they jump when misbehaving clients inflate the requests.15
  • Onion services hide the server too. The service keeps circuits open to a few introduction points and publishes a signed descriptor. A client picks a rendezvous relay, and both sides build circuits to it, so six relays sit between them and neither learns the other’s address. The service only makes outgoing connections and needs no open port.16
  • Bridges hide the fact of using Tor. They are unlisted entry relays, and pluggable transports such as obfs4 and Snowflake disguise the first hop for users on censored networks.17
identity readabledestination readablehiddenbothholds both endsISPVPN providerGuardExitWebsiteDirectbothyoudest.not on pathnot on pathnot on pathbothyoudest.VPNyoubothyoudest.not on pathnot on pathdest.Toryounot on pathyoudest.dest.VPN, then Toryouyoudest.dest.Tor, then VPNyoubothaccountdest.youdest.Tor, then VPN: the VPN provider sees the exit, not your address, but the paid account names you and every session leaves through it.
Who can link you to where you went, for five ways of routing the same visit. Each cell shows what one observer learns: your identity (magenta outline) and your destination (yellow); black means hidden. The magenta tag marks an observer that holds both ends on its own. Assumes HTTPS to forum.example.

The table puts the two halves of this chapter side by side. With one hop, some party always holds both ends. Tor spreads the ends across relays run by different people, so no relay is tagged. Chaining a VPN after Tor puts a fixed account at the end of every circuit, and that account was paid for by someone.

Where it gives way

Tor forwards cells as soon as they arrive. That keeps it fast enough for browsing, and it leaves a gap the Tor Project has been open about since the start: anyone who can watch traffic entering the network and traffic leaving it can match the two by timing and volume. The entry-guard design exists because of this, and the 2014 security advisory calls defeating such traffic confirmation “an open research problem”.1218 Watching both ends does not require owning relays. The RAPTOR study showed that a single autonomous system can see both ends of many circuits because internet routing is asymmetric, and it ran the correlation against live Tor traffic with 95% accuracy.19

Entry side, seen by the ISPsender address198.51.100.23destinationExit side, seen by forum.examplesender address192.0.2.201 (exit)destinationforum.examplebytes per moment0 s0.5 s1 s1.5 s2 sSame bursts, about 40 ms later on the exit side→ linkable by timingIllustrative traffic. Neither observer sees both addresses; together they do.
Traffic correlation. The ISP sees bursts leaving your address for a guard; the website sees the same bursts arriving from an exit a few tens of milliseconds later. Matching the two shapes links you to the site even though no relay saw both. This is what the home page figure means by “linkable by timing”.

Mixnets such as Loopix add random delays and cover traffic so that timing stops lining up, at the cost of seconds of latency; Tor accepted the opposite trade-off.20

The documented attacks on Tor users have mostly used relays the attacker ran:

  • From 30 January to 4 July 2014, a group of relays that had earned the Guard and HSDir flags inserted a signal into cells when onion-service descriptors were fetched or published, so that a colluding guard could tie a user or a service to an address.18 A federal judge later confirmed that an IP address in the Silk Road 2.0 related prosecution of Brian Farrell came from research by Carnegie Mellon’s Software Engineering Institute, obtained by FBI subpoena.21
  • The relay operator nicknamed KAX17 ran more than 900 relays at its peak, mostly guards and middles, which gave a user up to about a 16% chance of entering through one of them and about 35% of using one as a middle. Its purpose is unknown.2223
  • Another operator, tracked as BTCMITM20, ran thousands of malicious exit relays over time to replace Bitcoin wallet addresses in the web traffic passing through them.23 An exit can read and change whatever the application did not encrypt.
  • A local observer, such as a Wi-Fi operator, the ISP or the guard itself, can try to guess which site is loading from packet sizes and directions. The Deep Fingerprinting classifier reached 98.3% accuracy in a closed-world lab test, which overstates what works on the open web but shows the signal exists.24

In the cases that reached court, though, the network layer usually failed more plainly. Eldo Kim sent bomb threats to Harvard through Tor, but according to the FBI affidavit Harvard could see that he had connected to Tor over its own logged wireless network in the hours before the emails arrived.13 Tor hid what he sent and not that he was using Tor, there, then (case). Ardit Ferizi used no anonymising layer at all: the same Malaysian address that broke into a US retailer’s server had logged into his own Facebook and Twitter accounts (case).25 The Daily Beast reported, citing a single unnamed source, that the “Guccifer 2.0” persona once logged in without its VPN and left a Moscow address in a US company’s logs; that story is disputed and is not in the later indictment, which rests instead on shared servers and payments, covered in the infrastructure chapter (case).11

The other common route goes around Tor rather than through it. In the Freedom Hosting and Playpen investigations, the FBI took over the servers and delivered code to visitors’ browsers that sent the real address directly, outside the circuit (case). That is a failure of the device layer, and it is why a strong network layer does not protect a weak endpoint.

Common misconceptions

  • “A VPN makes you anonymous.” It moves the full view of your traffic from your ISP to one company. PureVPN’s and IPVanish’s records identified their customers.35
  • “Tor encrypts traffic all the way to the website.” Tor’s encryption ends at the exit. After that the traffic is protected only by whatever the application uses, such as TLS.23
  • “More hops mean more anonymity.” An observer who can see both ends does not care how many relays sit in between. A fixed hop after Tor can even reduce anonymity by tying sessions to one paid account.
  • “Tor has been broken cryptographically.” The documented deanonymisations used hostile relays, browser exploits or operator mistakes, not broken encryption.18
  • “The attack came from a domestic IP, so the attacker is domestic.” ORB networks and residential proxies place the last hop wherever the operator wants, often near the victim.10

Cases that failed here

Sources

Show all 25 sources
  1. ForbesTor users were caught by CMU researchers and lack 'reasonable expectation of privacy', judge rules
  2. Tor Project (support)Can I use a VPN with Tor?primary
  3. US Department of JusticeAffidavit in support of criminal complaint, United States v. Ryan S. Lin (D. Mass.)primary
  4. TorrentFreakPureVPN logs helped FBI net alleged cyberstalker
  5. TorrentFreakIPVanish 'no-logging' VPN led Homeland Security to Comcast user
  6. US Department of Justice911 S5 botnet dismantled and its administrator arrested in coordinated international operationprimary
  7. US Department of JusticeU.S. government disrupts botnet People's Republic of China used to conceal hacking of critical infrastructureprimary
  8. FBI, CNMF, NSA and partnersPeople's Republic of China-linked actors compromise routers and IoT devices for botnet operations (joint advisory)primary
  9. The Hacker NewsU.S. government disrupts Russian-linked botnet engaged in cyber espionage
  10. Mandiant (Google Cloud)IOC extinction? China-nexus cyber espionage actors use ORB networks to raise cost on defenders
  11. Tor Project (support)How does Tor work? Overviewprimary
  12. Tor Project (support)Entry guardsprimary
  13. Tor Project (specifications)Tor directory protocol, version 3: outlineprimary
  14. Tor MetricsRelays and bridges in the network (networksize.csv)primary
  15. Tor MetricsDirectly connecting users, all countries (userstats-relay-country.csv)primary
  16. Tor Project (community)Onion services: overviewprimary
  17. Tor Project (support)Unblocking Tor: bridges and pluggable transportsprimary
  18. Tor Project (blog)Tor security advisory: "relay early" traffic confirmation attackprimary
  19. USENIX Security Symposium (Sun et al.)RAPTOR: routing attacks on privacy in Torprimary
  20. USENIX Security Symposium (Piotrowska et al.)The Loopix anonymity systemprimary
  21. ThreatpostJudge confirms DoD-funded research was used to decloak Tor users
  22. nusenuIs "KAX17" performing de-anonymization attacks against Tor users?
  23. The RecordA mysterious threat actor is running hundreds of malicious Tor relays
  24. ACM CCS (Sirinam et al.)Deep Fingerprinting: undermining website fingerprinting defenses with deep learningprimary
  25. US District Court, E.D. VirginiaCriminal complaint and affidavit, United States v. Ardit Ferizi (E.D. Va., 1:15-mj-515)primary