← Breach analysis by the Dual Vulnerabilities framework

Marriott (Starwood) Guest Data Theft (Chinese MSS-linked actor, 2014)

2014 · United States · I - Accommodation and food service activities · Data theft via multi-year state-linked espionage intrusion

Analysed 2026-11-07 by Hermes

Open this case in the tool →

What happened

Data theft via multi-year state-linked espionage intrusion

Dual vulnerability type

Strategy to Execution

Evidence coverage: 4/6 stages, weight 32/60

Decisive stage: detection (clearest evidence) · Alternative: Expertise to Execution

Why this classification

(1) Expectation: Marriott carried Article 5(1)(f)/32 GDPR duties over the guest data on the acquired Starwood systems, ran a PCI-certified payment-card environment and a documented incident-response plan, had increased its IT-security investment before the intrusion was discovered and had decided and funded a migration away from Starwood's legacy estate; the ICO measured these duties against publicly available standards (NCSC 10 Steps, IAM guidance) and found Marriott had failed to heed that guidance in developing its cyber-security policies and procedures. (2) Practice: a web shell installed on 29 July 2014 (before the 2016 acquisition) enabled remote-access Trojans and Mimikatz credential harvesting, giving the attacker privileged, free movement inside the Starwood cardholder data environment; privileged and database activity went unmonitored (logging of CDE access not configured, alerts only on payment-card tables, no server logging of file creation), millions of passport numbers were stored unencrypted, whole guest tables were exported to dmp files in 2015-2016 unnoticed, and the four-year intrusion surfaced only through one Guardium alert relayed by Accenture on 7-8 September 2018; the ICO fined Marriott £18.4m in October 2020, finding four avoidable failings and negligence in maintaining systems with known shortcomings. (3) The knowing-doing gap is diagonal: formal strategic artefacts (regulatory obligations, PCI certification, a funded corporate security programme, a decided migration plan) never translated into the operational IT reality of the acquired legacy estate - the expectation was doctrinal and the break lay in execution, and no source shows leadership deciding against or deferring the named controls (budget rose before discovery; the migration slipped). The classification rests on the detection stage, whose origin and break are best documented (the ICO's first two failings plus the four-year dwell); the 2014 web-shell placement detail predates Marriott's ownership and is not load-bearing, and the ICO itself noted that the sophistication or specific vector of the attack is not the relevant focus. The response/disclosure stage shows no competing gap - the ICO dropped its Article 33 and 34 allegations and recorded prompt notification and full cooperation - and the entry/egress conditions point to the same doctrinal-to-execution diagonal, with Expertise to Execution the runner-up (see Alternative DV-Type). RM34 (Strategy to Execution lens): monitoring and alert-handling requirements were not operated as intended - alerts sat only on payment-card tables and privileged-account activity went unmonitored, so a four-year intrusion ended only on a single Guardium alert relayed by a third party (causal role: Obscuring). RM33 (Strategy to Execution lens): logging requirements were never implemented in the systems - no logging of access within the cardholder data environment and no server logging of file creation, so database exports went undetected and the scope of two deleted 2015/2016 table copies could never be established (causal role: Obscuring). RM26 (Strategy to Execution lens): the data-protection requirement was not implemented in storage - passport numbers were encrypted only inconsistently and without rationale, leaving about 5.25 million unencrypted and multiplying the harm of the theft (causal role: Amplifying). RM27 (Strategy to Execution lens): configuration baselines were not applied to deployed systems - the server hardening and whitelisting the ICO called readily available and mature solutions were absent, enabling privileged access, reconnaissance and traversal (causal role: Enabling).

Evidence by stage

Risk markers

Sources

Notes

The ICO Monetary Penalty Notice PDF itself (opened via web.archive.org snapshot of ico.org.uk/media/action-weve-taken/mpns/2618524/marriott-international-inc-mpn-20201030.pdf) is a 91-page scan with no text layer; pdftotext/pypdf return no text, so the MPN's detailed failings are taken from the ICO's own release (primary) and from two legal analyses quoting MPN paragraphs (Haynes Boone/shlegal.com, Clifford Chance, secondary). Attribution to the MSS rests on anonymous media statements (EuRepoC attribution_type: anonymous statement in media report; Wired notes there was no formal US-government attribution in December 2018, only a Secretary of State statement; El Pais reports the private-sector security-firm view); attribution is not load-bearing for the DV-Type. The initial compromise (29 July 2014) predates Marriott's September 2016 acquisition and the ICO's findings cover only 25 May - 17 September 2018, but the Commissioner held Marriott wholly responsible for that period and said due diligence is not a one-off requirement; Marriott also stated the Marriott and Starwood networks were always separate and only the Starwood network was affected (Senate testimony). Per the MPN analyses, the first 30 November 2018 notification emails omitted the call-centre telephone number and were re-sent on 9 December 2018 - a minor disclosure imperfection the ICO did not sanction. Class actions continued after the fine (Fourth Circuit 2023 remand on class-action waiver, EFF blog). Row name 'Marriott Hack' matches the row description (approx. 500 million guests, Chinese intelligence-gathering also hitting health insurers and clearance files); no data-quality mismatch found.