why hackers don’t get caught

Behavior: habits and people

Working hours, holidays, language and the people someone trusts all keep a pattern, and a pattern can be read without breaking any technical layer.

What it hides
Habits, hours, language, the way someone talks
How it gives way
One careless sentence or an activity clock
How much to read

The technical layers can all hold and the person can still be found, because a person has a routine. They work at certain hours, take certain holidays, write in a certain language, talk to certain people and trust some of them. This layer is about that routine, and about the people around an operation, who know things no protocol can hide.

The clock

Every logged event has a timestamp, and a timestamp is a time of day somewhere. When an analyst collects enough events from one actor, the shape of a working week appears, and the only question is which time zone makes it look normal.

Compile times

Malware carries its own clock. A Windows executable records the time its author compiled it. FireEye’s 2014 report on APT28 found that more than 96% of the group’s samples were compiled on weekdays, and 89% between 08:00 and 18:00 in UTC+41. FireEye summarised this as developers working business hours “consistent with the time zone of Russia’s major cities, including Moscow and St. Petersburg”2.

Bitdefender, working from its own collection of APT28 files a year later, spelled out the method. It counted, for every time zone, how many binaries had been compiled between 08:00 and 18:00 on a weekday, and looked for the zone that stood out. UTC+4 did, with 88% of the files3. Two vendors with different samples arrived at the same band.

The same week of logged activity, read in three clocks0003060912151821UTCstarts at 04:00, stops at 14:00MonTueWedThuFriSatSunUTC+4an office weekMonTueWedThuFriSatSunUTC+8starts at noon, runs to 22:00MonTueWedThuFriSatSunframe: 08:00–18:00, Mon–Friactive hourIllustrative data, shaped like FireEye's published APT28 compile-time statistics.
Illustrative data, shaped like the APT28 compile-time statistics. The same log is drawn three times, read in three clocks, which is the test Bitdefender describes run on three candidate zones. Only in UTC+4 does the activity fill the 08:00–18:00 weekday frame.

The shape of a week

The shape carries more than an offset. PwC and BAE Systems lined up three independent clocks for APT10: when its domains were registered, when its PlugX, RedLeaves and Quasar samples were compiled, and when its operators worked hands-on inside victim networks. Shifted to UTC+8, the registrations fell into Chinese business hours with a two-hour midday break, and the hands-on activity ran mainly from midnight to 10:00 UTC, which is 08:00 to 18:00 in UTC+84. In December 2018 the U.S. indicted two men it said had worked for APT10 through a Tianjin company, in association with the Tianjin bureau of China’s Ministry of State Security5. The clock had pointed at UTC+8; the indictment named a city in it.

A week can also start on the wrong day. FireEye placed APT33’s activity in a time zone about four and a half hours ahead of UTC, which matched Iran’s daylight time, and found it ran largely from Saturday to Wednesday, with no activity on Thursday. That is Iran’s working week6. Kaspersky read the opposite side of the world from the Equation group’s timestamps: working days and hours of “a well-organized software developer” in UTC−3 or UTC−4, if the day starts at 8 or 97.

Criminal groups look different. Group-IB, analysing Conti’s attacks in 2022, found the group active about 14 hours a day, starting near noon in GMT+3 and slowing only after 21:00, with weekends worked and only the New Year holidays off8. A large operation has shifts and roles, and its clock is the sum of many people’s clocks.

Five published working weeks, each in the local time its report inferreddayshours, localMTWTFSS0006121824APT28FireEye 2014, Bitdefender 2015UTC+4an office weekAPT10PwC and BAE 2017UTC+8not statedtwo-hour midday breakAPT33FireEye 2017UTC+4:30not statedSaturday to WednesdayEquationKaspersky 2015UTC−3 or −4not statedworkdays; starts 8 or 9ContiGroup-IB 2022GMT+314 hours, weekends tooactivedeclining
What five reports published, each drawn in the local time the report inferred. Blank strips are things the report did not state. The time zone comes from the shift that makes the week ordinary; the shape itself (a lunch break, a Saturday start, seven-day shifts) says what kind of employer keeps it.

Holidays and gaps

A working week also has gaps that belong to a calendar. Proofpoint observed that one of the main distributors of the Dridex banking malware was completely inactive from 24 December 2015 through 10 January 2016, with test runs starting again on 11 January and volume back to normal by the 13th. Proofpoint linked the shape of that break to the Orthodox calendar, under which Christmas falls on 7 January and the week before it is celebrated as a New Year holiday9. A Western defender expecting a quiet Christmas Day would have seen nothing unusual; the gap that matters is the one that ends on the eleventh.

What the clock cannot say

Reading a clock this way has known pitfalls. Daylight saving time moves a working day by an hour twice a year, so data spanning months smears the edges. Timestamps can come from the attacker’s machine, a compromised server in another country, or a build system set to UTC, and each tells a different story. What survives these checks is usually a band of two or three neighbouring time zones, not a city.

Timestamps are also easy to edit. Bitdefender found APT28 files compiled after 2013 whose file creation times had been set to 14 April 2008, 16:00, so that a victim searching for recently added files would pass over them3. And an operator can choose hours for the victim’s sake rather than their own: Kaspersky found the Lazarus group working inside one bank outside the victim’s office hours, in the victim’s time zone, to avoid being noticed10. Hands-on intrusion times then describe the target’s schedule as much as the attacker’s.

So a clock is strongest when another layer agrees with it. Mandiant’s 2013 report on APT1 traced the operators’ own remote desktop connections to their attack infrastructure: of 832 source IP addresses, 817 belonged to networks in China11. The clock suggests a time zone; the addresses say which country in it.

Language

Software records more of its author’s language than the author intends. In the same APT1 investigation, 1,849 of the 1,905 remote desktop sessions Mandiant observed carried the keyboard layout “Chinese (Simplified) – US Keyboard”, a value the Remote Desktop client fills in from the operator’s own language settings11. FireEye found APT28 samples built with Russian language settings since 20071, and Bitdefender found an APT28 tool whose debug path ran through a folder called “Пользователь”, Russian for “user”3. FireEye found Farsi artefacts in two of APT33’s tools6.

These artefacts are cheap to plant, and analysts know it. Kaspersky found Russian words scattered through exploits and a backdoor used by Lazarus, and judged them “a very cheap imitation, because every native Russian speaking software developer quickly noticed how odd these commands were”10. When the U.S. charged six GRU officers over the Olympic Destroyer attack on the 2018 Winter Olympics, the Justice Department said the conspirators “took steps to pin the Olympic Destroyer attack on North Korea”12. The forged build header that gave it away is covered in the attribution chapter3.

A language sample says what the writer can write, not where they hold a passport. Flashpoint compared the 28 ransom notes shipped with WannaCry and found that most were machine translations of the English one, that the Chinese notes read as written by someone fluent, and that the English note had a glaring grammatical error13. The U.S. later charged a North Korean programmer, Park Jin Hyok, over WannaCry, and described his employer as a front company with offices in China and North Korea3. Both findings can be true at once.

Guccifer 2.0 is the clearest case of a persona’s language being traced to where it was made. The persona claimed to be a lone Romanian hacker28, and Romanian speakers who read its interview with Motherboard judged that it was not a native speaker14. The 2018 indictment went further. On 15 June 2016, between 16:19 and 16:56 Moscow time, the conspirators searched for English phrases such as “some hundred sheets” and “worldwide known” on a Moscow server run by GRU Unit 74455; at 19:02 the first Guccifer 2.0 post appeared, using those phrases28. The persona’s English was assembled where it was written, and a search log kept the drafts. See the Guccifer 2.0 case.

Time against a life

The strongest use of timing is a match against one person’s known day: was every moment the pseudonym acted a moment the candidate could have acted?

Eldo Kim’s bomb threats reached Harvard at about 8:30 in the morning, half an hour before a final exam he was due to sit61. The FBI affidavit says Harvard was able to determine that “in the several hours leading up to the receipt of the e-mail messages”, Kim had accessed Tor using Harvard’s wireless network15. Tor hid what he did, not that he was using Tor on a logged network at the time that mattered. Josephine Wolff pointed out the limit: if many people on the network had been using Tor, the list would not have narrowed, and the match alone did not prove he sent the messages15. He confessed when interviewed61. See the Eldo Kim case.

Jeremy Hammond was matched over several days. Agents identified the wireless router in his Chicago building, obtained a court order to record its addressing information, and could see his laptop connecting to Tor. The laptop was active when agents saw Hammond at home and when he was chatting with Monsegur, the informant, and the activity stopped or dropped whenever he left the building16.

One alias's online sessions checked against two candidates' daysday 1day 2day 3day 4Candidate ACandidate Balias onlineinformant's chat listA at homeagents watchingTor from A's linepen registeralias onlineinformant's chat listB at homeagents watchingevery session falls inside A's home hours and A's Tor trafficonline while B is out: B is ruled outIllustrative data, shaped like the reported surveillance of Jeremy Hammond, March 2012.
Illustrative data, shaped like the reported surveillance of Jeremy Hammond. The alias's sessions are identical in both panels. Each session that falls inside a candidate's home hours is weak on its own; six in a row is not. One session outside them is enough to rule a candidate out.

The same arithmetic undoes an alibi built from timing. On 16 May 2014, while the defendant in the Japanese remote-control virus case sat in court for the eighth hearing of his trial, an email claiming to be from the “real culprit” went out at 11:3717. Investigators suspected he had set a smartphone timer, and police had already watched him acting suspiciously on a Tokyo riverbank before the email arrived57. The timer made the clock say he was elsewhere; the surveillance made it say where the phone was. See the remote-control virus case.

The single slip

The popular story of anonymity failing is a single slip: careful for months, careless once. Side by side, the cases that look like that share less than the story suggests.

Five single-slip stories: the cover, the moment, who recorded it, how solid the account iscasecover, most of the timethe momentrecorded byaccountSabu2011proxiesone IRC login from homeIRC serverlogsdisputedPC remote2013others' infected PCswent to Enoshima in personsecuritycamerascourt recordEldo Kim2013Tor, throwaway mailTor from campus Wi-FiHarvard'snetwork logsaffidavitGuccifer 2.02016VPNone logon without VPNa US socialmedia companyone sourceBluenoroff2017proxies and VPNsone connection from NKits own C2server in Europevendor forensics
Five single-slip accounts. The black bar is the cover that held almost all the time, from the point of view of whoever kept the log; the magenta mark is the one moment it did not. A slip mattered only where someone recorded it and kept the record. The last column is how solid each account is.

The best-known one is disputed. Fox News, which broke the story of Hector Monsegur’s arrest, reported that Sabu of LulzSec normally hid behind proxies but once logged into an IRC server from his own IP address7. Monsegur has rejected that account and says he was identified from an older dox about him8. The site lists the case as disputed. See the Sabu case.

Guccifer 2.0’s reported slip rests on one source. The Daily Beast, citing a person familiar with the investigation, reported that on one occasion the operator failed to turn on the VPN before logging in, leaving a Moscow IP address in an American social media company’s logs11. The indictment does not rest on that; it uses the server searches above and a shared pool of bitcoin that paid for the persona’s infrastructure28.

A vendor investigation gives the cleanest example, because it shows the log as well as the mistake. On a command server in Europe used by Lazarus’s Bluenoroff unit, the operator had connected through proxies and VPN servers from countries ranging from France to Korea, but “one short connection was made from a very unusual IP range, which originates in North Korea”. The server’s logs had survived only because cryptocurrency-mining software installed on it froze the system before it could be cleaned18. Inside one bank, the same group enabled a system monitoring driver to debug a problem and then forgot to wipe its event log, which recorded the commands it ran10.

The pattern across these is pressure and a record. Kim was half an hour from an exam. The Lazarus operators were fixing crashes on a live system. In each case something outside the actor’s control (a campus network, a social media company, a frozen server, a camera on Enoshima) wrote the moment down and kept it until an investigator asked56.

People: informants, insiders and undercover officers

The most reliable way into a group is often a member of it. This is the people failure in the case files: no protocol or habit failed, someone who knew the truth told it.

The carding forum DarkMarket was believed to be run by a member called “Master Splyntr”. In 2008 German public radio reported that Master Splyntr was an FBI agent, J. Keith Mularski, and the FBI said the sting, which had run for about two years, had produced 56 arrests19. The traders had been dealing on a forum run by the investigator.

An informant can work from inside a group’s conversations. After his arrest in June 2011, Monsegur cooperated with the FBI and kept running his Sabu persona under its direction; prosecutors credited him with helping identify other members of LulzSec and AntiSec9. Hammond discussed the Stratfor hack in chats with Sabu, and mentioned past arrests and time in federal prison; the FBI matched those details against arrest records to narrow the suspects to him10. The surveillance described above leaned on the same informant: the laptop’s activity was read against Hammond’s chats with Monsegur16. No protocol can tell a loyal partner from one who has already been arrested. See the Hammond case.

Remote work that hides a country

A behaviour pattern can also be manufactured. North Korean IT workers have been placed in remote jobs at U.S. companies using stolen or borrowed U.S. identities. The company ships its laptop to a home in the United States; the person living there keeps it running and connects it to remote-access hardware such as KVM switches, and the worker abroad operates it from overseas20. To the employer, the worker appears to connect from a U.S. residential address52.

A laptop farm, as seen by the employer and by an investigatorWorkerabroadlocation hiddenremoteA home in the USlaptop issued toidentity AEmployer 1US IPUS addresslaptop issued toidentity BEmployer 2US IPUS addresslaptop issued toidentity CEmployer 3US IPUS addressEach employer's viewInvestigator's view: one house, many identities.
A laptop farm. Each employer sees a U.S. IP address and a U.S. shipping address. The worker abroad stays out of view, but every identity in the scheme now runs through one house.

The arrangement hides the worker’s location by concentrating many personas in one place. Christina Chapman, sentenced in July 2025 to 102 months, hosted such a farm at her home in Arizona: the scheme used 68 stolen identities, placed workers at 309 U.S. companies and generated more than $17 million53. In June 2025 the Justice Department announced searches of 29 known or suspected laptop farms across 16 states20. See the laptop farm case.

The rest of the routine still shows. The FBI’s guidance to employers lists behavioural red flags: reluctance to appear on camera, inconsistencies in time, location or appearance when they do, and language preferences set to Korean by someone who claims to be from a country where Korean is not spoken21. A later warning adds multiple logins into one account within a short time from IP addresses in different countries3. The laptop sits in the United States; the clock and the language settings of the person using it do not.

Where it gives way

The behaviour layer rarely fails in a single moment. It fails through accumulation: enough timestamps to show a clock, enough text to show a language, enough sessions to match one person’s day, enough trust for an informant to work. Most cases on this site have a behavioural element, but the decisive one is usually elsewhere.

The Japanese remote-control virus case of 2012 is the exception that shows the layer on its own. The culprit posted threats from other people’s infected computers, and police first arrested four innocent owners because they trusted the IP addresses. What identified him was physical behaviour: a puzzle email of his own led to a memory card on a cat’s collar on Enoshima, where cameras had recorded him56, and on bail he was watched burying the smartphone that held traces of the fake “real culprit” email57. The infected computers had hidden him. His own movements, timed and filmed, did not.

Common misconceptions

  • “Attackers can be located by their working hours alone.” Timing gives a band of candidate time zones, and timestamps can be edited. Analysts weigh it with infrastructure and language evidence.
  • “Criminal groups keep office hours.” Some actors look like it. Conti’s measured activity ran about 14 hours a day, weekends included.
  • “The language in the code tells you the author’s nationality.” It tells you what the writer could write. Lazarus planted Russian words, and the fluent Chinese in WannaCry’s notes sits alongside U.S. charges against a North Korean.
  • “Most anonymous actors are caught by one slip.” The famous single-slip stories are often disputed or rest on one source. Where a slip did matter, it mattered because a log kept it and other evidence pointed the same way.
  • “Careful technology protects against informants.” Hammond’s chats went to Sabu, who was already working for the FBI. No encryption changes who is reading on the other end.
  • “A U.S. IP and a U.S. address mean a U.S. worker.” For laptop farms, both belong to the facilitator’s house.

Cases that failed here

Sources

Show all 21 sources
  1. ThreatpostRussian APT28 Group Linked to NATO, Political Attacks
  2. FireEye (Mandiant), republished on the Google Cloud Threat Intelligence blogAPT28: A Window Into Russia's Cyber Espionage Operations?
  3. BitdefenderAPT28 Under the Scope: A Journey into Exfiltrating Intelligence and Government Information
  4. PwC UK and BAE SystemsOperation Cloud Hopper: Exposing a systematic hacking operation with an unprecedented web of global victims
  5. U.S. Attorney's Office, Southern District of New YorkTwo Chinese Hackers Associated With The Ministry Of State Security Charged With Global Computer Intrusion Campaigns Targeting Intellectual Property And Confidential Business Informationprimary
  6. FireEye (Mandiant), republished on the Google Cloud Threat Intelligence blogInsights into Iranian Cyber Espionage: APT33 Targets Aerospace and Energy Sectors and has Ties to Destructive Malware
  7. Kaspersky GReAT (Securelist)Inside the EquationDrug Espionage Platform
  8. Group-IBThe Conti Enterprise: ransomware gang that published data belonging to 850 companies
  9. ProofpointAn Orthodox Holiday For Some Unorthodox Actors
  10. Kaspersky LabLazarus Under the Hood
  11. MandiantAPT1: Exposing One of China's Cyber Espionage Units (copy at the National Security Archive)
  12. U.S. Department of Justice, Office of Public AffairsRemarks by Assistant Attorney General for National Security John C. Demers on Announcement of Charges Against Russian Military Intelligence Officersprimary
  13. BleepingComputer (reporting Flashpoint's analysis)Linguistic Analysis Suggests WannaCry Ransomware Is the Work of a Chinese-Speaking Crook
  14. Motherboard (Vice)Why Does DNC Hacker 'Guccifer 2.0' Talk Like This?
  15. Slate (Josephine Wolff, quoting the FBI affidavit)How the Alleged Harvard Bomb Hoaxer Failed to Cover His Online Tracks
  16. NextgovFBI nets cyber informants with hacktivist sting
  17. IWJ Independent Web JournalPC遠隔操作事件「真犯人」を名乗る人物からメール 片山氏「これをもって、この裁判を終わりにしてほしい」
  18. Kaspersky GReAT (Securelist)Lazarus Under The Hood
  19. CSO OnlineFBI Says Dark Market Sting Netted 56 Arrests
  20. U.S. Department of Justice, Office of Public AffairsJustice Department Announces Coordinated, Nationwide Actions to Combat North Korean Remote Information Technology Workers’ Illicit Revenue Generation Schemesprimary
  21. FBI, Internet Crime Complaint Center (IC3)Additional Guidance on the Democratic People's Republic of Korea Information Technology Workersprimary