Attribution answers "who did this", but security teams regularly misread it: they treat a vendor's cluster name as a confirmed state actor, mix up aliases, or let a public attribution drive decisions that the evidence does not support. Governments, intelligence agencies and security vendors attribute cyber operations in different ways, for different purposes, and with carefully chosen confidence language. Knowing how that process works helps you judge how much weight to give an attribution, how to brief leadership, and when attribution matters at all for your response. This guide explains the levels of attribution, the evidence used, how vendors name and merge actors, how official statements express confidence, and how to use attribution in your own threat intelligence and incident work. It is for threat intelligence analysts, incident responders and security leaders. It builds on the Threat Intelligence Guide.
Key takeaways
- Attribution has levels. Linking activity to a tracked cluster, to a country, to a specific agency or unit, and to named individuals are different claims that need different evidence.
- Confidence and likelihood are separate. US intelligence standard ICD 203 treats confidence in a judgment and the probability of an event as distinct, and official statements are worded with that in mind.
- Vendor names are working labels. Mandiant UNC clusters and Microsoft Storm-#### designations are temporary until the vendor reaches high confidence, and clusters merge or split as evidence grows.
- Public attribution is a political act. Governments decide when and whether to name a state, often jointly with allies and alongside sanctions or indictments, so timing tells you about policy as well as evidence.
- Defenders rarely need a name to act. Behaviour and techniques drive detection; attribution matters most for risk assessment, legal and insurance questions, and leadership communication.
What attribution means: the levels
| Level | Claim | Typical evidence | Who usually makes it |
|---|---|---|---|
| Technical clustering | These intrusions were carried out by the same operator | Shared infrastructure, malware code, tooling, tradecraft and timing | Vendors, CERTs, in-house CTI |
| Named group | The cluster is a known, tracked actor such as APT29 or Volt Typhoon | Accumulated overlap with a previously defined group over years | Vendors, ATT&CK (which records groups and aliases from public reporting) |
| State nexus | The actor works for or on behalf of a country | Targeting aligned with state interests, working hours, language artefacts, plus intelligence sources | Vendors (assessed), governments |
| Specific organisation | A named agency or unit is responsible, for example GRU Unit 74455 | Intelligence collection, signals and human sources not available to the private sector | Governments and intelligence agencies |
| Individuals | Named people carried out the activity | Law enforcement investigation, financial records, legal process | Prosecutors, courts, sanctions authorities |
Each step up the table needs a different kind of evidence. A vendor can reach high confidence at the clustering level with telemetry alone, but naming a government unit or a person usually relies on intelligence or legal powers. When reading a report, find which level it is claiming.
Evidence used in attribution
- Tactical: malware, domains, IP addresses and exploited CVEs. Mandiant describes this as the level that feeds block lists and signatures. It is the easiest to share and the easiest for an adversary to change or fake.
- Operational: behaviour patterns such as tools used, operation frequency, target sectors and locations. In ATT&CK terms this is the technique profile of a group, which is harder to change.
- Strategic: motives, goals, sponsors and associates. This is where state attribution sits, and where private sector evidence is thinnest.
- Intelligence and legal sources: governments draw on intercepted communications, human sources, financial records and court process. The 2018 US indictment of twelve GRU officers, for example, describes the defendants searching online for English phrases that later appeared in the Guccifer 2.0 persona's first blog post.
- Analytic models: the Diamond Model (adversary, capability, infrastructure, victim) and ATT&CK help structure the comparison between intrusions. See the Threat Intelligence Guide for these models.
Deception and false flags
Adversaries plant misleading evidence. The UK government said in October 2020 that the GRU's cyber unit "attempted to disguise itself as North Korean and Chinese hackers" when it targeted the 2018 Pyeongchang Winter Olympics opening ceremony. The 2016 Guccifer 2.0 persona falsely claimed to be a lone Romanian hacker, according to the 2018 indictment. Single indicators such as a language string, a compile time or a reused tool are weak evidence on their own because they are easy to fake or share.
How vendors name and track actors
| Vendor scheme | Working label | Mature label | Notes from the vendor |
|---|---|---|---|
| Mandiant (Google) | UNC#### (uncategorised), TEMP.Name | APT## (state), FIN## (financial) | A UNC is built around an anchoring characteristic, can merge or split, and may graduate after years of analysis, as UNC902 became FIN11 |
| Microsoft | Storm-#### (developing group) | Weather family name: Typhoon (China), Blizzard (Russia), Sandstorm (Iran), Sleet (North Korea); Tempest (financially motivated); Tsunami (private sector offensive actors); Flood (influence operations) | Storm designations are temporary "until high confidence is reached about the origin or identity of the actor" |
| MITRE ATT&CK | Not applicable | Gxxxx group IDs with associated names | Records aliases from public reporting, for example G0016 APT29 and G1055 VOID MANTICORE, without making its own attribution |
Because each vendor sees different telemetry, their clusters do not line up exactly. One vendor's group may be two groups for another. Microsoft publishes an alias mapping on GitHub with a KQL function for Defender XDR, which helps with translation but does not guarantee that two names describe the same activity. When you map aliases, record the source and the overlap stated (for example "overlaps with", "associated with" or "same as"). The NHI Mgmt Group actor profiles, such as the APT29 (Midnight Blizzard) Threat Actor Profile and the Volt Typhoon Threat Actor Profile, list aliases with their sources for this reason.
Confidence levels and estimative language
ICD 203, the US Intelligence Community analytic standards directive, requires products to express uncertainty properly and to distinguish intelligence information from assumptions and judgments. It sets terms for likelihood and treats confidence separately.
| ICD 203 likelihood term | Approximate probability |
|---|---|
| Almost no chance, remote | 01 to 05% |
| Very unlikely, highly improbable | 05 to 20% |
| Unlikely, improbable | 20 to 45% |
| Roughly even chance, roughly even odds | 45 to 55% |
| Likely, probable | 55 to 80% |
| Very likely, highly probable | 80 to 95% |
| Almost certain, nearly certain | 95 to 99% |
- Confidence: ICD 203 describes confidence as resting on the logic and evidence behind a judgment, including the quantity and quality of sources. It says a confidence level and a degree of likelihood should not be combined in the same sentence.
- UK practice: UK statements also use calibrated terms. The NCSC assessed in 2020 "with high confidence" that Olympic-related attacks were "almost certainly (95%+)" carried out by GRU unit GTsST, and in 2018 that the Russian military was "almost certainly responsible" for NotPetya.
- Read the wording closely: the UK's December 2023 release on Star Blizzard described the group as "almost certainly subordinate" to FSB Centre 18 in one place and "highly likely subordinate" in another. Quote the exact phrase from the primary source when you brief.
- Vendor wording: vendors use "assess with moderate confidence", "likely" and "overlaps with" more loosely. Treat a vendor's state attribution as an assessment, not a finding of fact, unless a government has confirmed it.
Public attribution by governments
| Statement | Date | Attribution as stated | Accompanying action |
|---|---|---|---|
| UK on NotPetya | 15 February 2018 | Russian military; NCSC "almost certainly responsible" | Public condemnation |
| US indictment, 2016 election | 13 July 2018 | Twelve GRU officers in Units 26165 and 74455 | Criminal charges |
| UK on Olympic Games attacks | 19 October 2020 | GRU GTsST (Unit 74455, Sandworm); high confidence, almost certainly (95%+) | Joint exposure with US indictments |
| US on SolarWinds | 15 April 2021 | Russian SVR (APT29, Cozy Bear); Intelligence Community "high confidence" | Sanctions on six Russian technology companies, diplomatic expulsions |
| Volt Typhoon advisory AA24-038A | 7 February 2024 | "PRC state-sponsored cyber actors"; high confidence of pre-positioning | Joint advisory with Australia, Canada, the UK and New Zealand |
Three patterns stand out. First, statements usually follow incidents by months or years, because governments protect sources and co-ordinate with allies. Second, attribution increasingly comes with costs imposed: sanctions, indictments or, in the EU, listings under the cyber sanctions framework established by Council Decision (CFSP) 2019/797, adopted on 17 May 2019. Third, joint statements by several countries signal political consensus as well as shared evidence. The SolarWinds case is covered on the SolarWinds Supply Chain Compromise breach page.
Using attribution in your own programme
When attribution matters
- Risk and threat modelling: knowing which state or criminal groups target your sector shapes priorities. See the Threat Actor Groups Guide.
- Response decisions: a state-linked destructive actor or a ransomware affiliate implies different containment and recovery urgency and different law enforcement contacts.
- Legal, regulatory and insurance: sanctions rules can affect ransom payment decisions, and some insurance policies contain war or state-action exclusions, so wording in your records matters. Involve legal counsel before writing attribution into formal documents.
- Leadership and public communication: executives will be asked "who was it". Give the sources' wording rather than your own stronger version.
When it does not
Detection engineering, containment and hardening rarely depend on a name. ATT&CK techniques and observed behaviour are enough to build detections, and many actors share the same tools. Avoid delaying containment while attribution is debated.
How to implement an attribution practice
- Define levels and wording: adopt a house scale based on ICD 203 likelihood terms and three confidence levels (low, moderate, high), and state which attribution level each claim makes.
- Cluster before you name: track your own intrusions as internal clusters with anchoring evidence, and only map them to public group names when overlap is substantial.
- Record aliases with sources: keep an alias table that records who uses each name and the stated relationship.
- Separate facts from judgments: in reports, list observed evidence first, then the assessment, then alternatives considered, as ICD 203 requires of US analysts.
- Review and revise: revisit attributions when new reporting appears and explain changes, so consumers can see why an assessment moved.
- Test with red cells: for high-impact attributions, ask a colleague to argue the alternative, including a false flag.
Decision guidance by organisation size
- Small teams: consume attribution from government advisories and vendor reports; do not attempt your own state attribution. Focus on the techniques described.
- Mature CTI teams: maintain internal clusters, share indicators with ISACs and national CERTs, and contribute evidence rather than names.
- Regulated and critical sectors: establish a channel to the national authority (NCSC, CISA, FBI) so that your evidence reaches those who can make state-level attributions.
Reading an attribution claim: five questions
- Who is making the claim? A government statement, an indictment, a vendor report and a persona's own boast carry very different weight. Prefer primary sources over news summaries, which often drop qualifiers.
- Which level is claimed? "Consistent with Chinese state interests" is a state-nexus assessment; "Unit 74455" is an organisational attribution that private sector telemetry alone rarely supports.
- What confidence and likelihood terms are used? Copy them verbatim. If none are given, say so.
- What evidence is disclosed? Technical overlaps you can check yourself are more useful than unexplained assertions, but governments often withhold evidence to protect sources.
- What alternatives were considered? Good reports discuss shared tooling, contractors and false flags. If a report does not, treat its conclusion more cautiously.
Applying these questions consistently keeps briefings accurate and stops one vendor's naming choices from reshaping your whole threat model.
Where identity fits
Identity telemetry is some of the strongest attribution evidence a defender holds. Sign-in patterns, user agents, authentication methods, OAuth application registrations and the way an actor uses service principals and tokens are behavioural fingerprints that persist across infrastructure changes. Storm-0558's use of a forged token-signing key, for example, became a defining characteristic of that cluster. Preserve identity provider logs for long enough to support later attribution. See the Storm-0558 and Cloud Key Theft Profile and the Threat Intelligence for Identity and NHI Guide.
Detection and measurement
- Alias accuracy: percentage of actor names in your threat library with a cited source and stated relationship.
- Wording compliance: share of CTI reports that use the house likelihood and confidence scale correctly.
- Revision rate: number of attributions changed after review, with reasons recorded (a healthy programme revises).
- Behaviour coverage: percentage of techniques attributed to your priority actors that have a tested detection, which shows attribution turning into defence.
- Time to action: time from a government advisory naming an actor relevant to you to completed hunts for its techniques.
Framework mapping
| Activity | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 |
|---|---|---|
| Receiving and analysing threat intelligence | ID.RA-02, ID.RA-03 | PM-16 Threat Awareness Program, RA-3 |
| Analysing incidents and their cause | RS.AN-03 | IR-4 |
| Sharing information with partners | RS.CO-03 | PM-15, SI-5 |
| Preserving evidence and logs | RS.AN-07 | AU-11 |
Common failure patterns
- Turning "overlaps with" into "is": merging vendor clusters that only share a tool produces a profile that fits nobody.
- Upgrading the confidence: repeating "moderate confidence" as "confirmed" in an executive brief or press statement.
- Trusting a single indicator: a language string or reused tool is easy to fake, as the Pyeongchang false flags showed.
- Believing the persona: hacktivist and criminal personas are sometimes state fronts, as with Guccifer 2.0 and Handala. See the Hack and Leak Operations Guide.
- Delaying response for a name: containment should proceed on behaviour while attribution develops.
- Losing the evidence: short log retention makes later attribution, legal action and insurance claims impossible.
Practitioner checklist
- Adopt a written likelihood and confidence scale based on ICD 203 and use it in every report.
- State the attribution level (cluster, group, state, organisation, individual) for each claim.
- Maintain an alias table with source, date and stated relationship for each name.
- Quote government and vendor attribution wording exactly when briefing leadership.
- Prioritise detections by the techniques of actors relevant to your sector, not by names alone.
- Retain identity, endpoint and network logs long enough to support later attribution.
- Agree with legal counsel how attribution is recorded in incident reports and insurance claims.
- Share evidence with national authorities and ISACs rather than publishing your own state attributions.
Standards and references
- US Office of the Director of National Intelligence: Intelligence Community Directive 203, Analytic Standards
- UK Government: UK Exposes Series of Russian Cyber Attacks Against Olympic and Paralympic Games
- UK Government: Foreign Office Minister Condemns Russia for NotPetya Attacks
- The White House: Fact Sheet, Imposing Costs for Harmful Foreign Activities by the Russian Government
- CISA: PRC State-Sponsored Actors Compromise and Maintain Persistent Access to US Critical Infrastructure (AA24-038A)
- Microsoft Learn: How Microsoft Names Threat Actors
- Google Cloud (Mandiant): How Mandiant Tracks Uncategorized Threat Actors
- EUR-Lex: Council Decision (CFSP) 2019/797 on Restrictive Measures Against Cyber-Attacks
Related NHI Mgmt Group resources: Threat Intelligence Guide · Threat Actor Groups Guide · APT29 (Midnight Blizzard) Threat Actor Profile · Hack and Leak Operations Guide