What is it?
Golden SAML is an attack pattern in which a stolen federation signing key is used to forge trusted SAML assertions. The learning goal is to identify the trust boundary being abused, the attacker’s sequence, and the telemetry that records it.
How it works in plain English
Use this three-part model to understand the attack without memorizing a tool name.
The system accepts a credential, token, ticket, recovery decision, or identity assertion as proof of who the requester is.
The attacker obtains, guesses, forges, replays, or changes that authentication material.
Focus on how the identity was acquired, where it was used, what device and network context changed, and what happened immediately after success.
Worked example: how the sequence may look
Access to or influence over the relevant Identity Provider, SaaS environment.
Trigger the condition that enables golden saml.
Authentication failures and successes, ticket or token activity, MFA events, and account changes.
Baseline the affected identity, host, application, service, or protocol.
If successful, the attacker may continue with Cloud Account Takeover. This is a possibility, not a guaranteed next step.
This is a defensive learning scenario. Real incidents vary, and the listed outcome is only one possible path.
Why do attackers use it?
Advance an objective associated with credential access.
Exploit a weak control, unsafe default, exposed service, trusted relationship, or human decision.
Create access, gain privilege, evade defenses, steal information, or disrupt operations.
Prerequisites
Access to or influence over the relevant Identity Provider, SaaS environment.
Knowledge of the target’s technology, identities, configuration, or user behavior.
A weakness, misconfiguration, exposed interface, stolen secret, or social opportunity.
A path to observe success and continue toward the objective.
Attack path possibilities
Attack chains are not fixed. These links show common possibilities to investigate before and after this behavior.
A targeted message links to credential theft or malware delivery.
T1110.001 Password GuessingCommon or targeted passwords are tested against one account.
T1003.001 LSASS Credential DumpingCredential material is extracted from the Windows security process.
T1557 Adversary-in-the-Middle PhishingA reverse proxy captures credentials, MFA exchanges, and session cookies in real time.
Stolen credentials, tokens, or recovery paths control a cloud identity.
T1021 Remote ServicesRDP, SMB, SSH, WinRM, VNC, or similar services are used to access remote systems.
T1550.002 Pass-the-HashA stolen password hash is used directly for authentication without recovering the plaintext.
T1550.003 Pass-the-TicketA stolen Kerberos ticket is reused to access resources without the account password.
Step-by-step attack flow
Reconnoiter the target and identify the exposed trust boundary.
Prepare the infrastructure, lure, input, credential, or payload required.
Trigger the condition that enables golden saml.
Confirm access or effect while attempting to avoid controls.
Use the result for follow-on access, execution, collection, movement, persistence, or impact.
Remove evidence, rotate infrastructure, or repeat against other targets.
MITRE ATT&CK mapping
ATT&CK is updated over time. Verify the live technique before operationalizing a detection.
Logs and artifacts
Authentication failures and successes, ticket or token activity, MFA events, and account changes.
Identity-provider risk, device, source-network, conditional-access, and session telemetry.
Endpoint evidence of credential access, browser data access, or security-process interaction.
Realistic log examples
These records use realistic field names and formats but synthetic organizations, users, addresses, and identifiers. A single event is not proof; correlate time, identity, source, target, and the resulting action.
Kerberos service-ticket request; useful for roasting, ticket abuse, and unusual service access.
2026-07-16T14:25:10Z EventID=4769 Computer=DC01.corp.example
TargetUserName=analyst@CORP.EXAMPLE ServiceName=MSSQLSvc/sql01.corp.example:1433
IpAddress=10.24.18.77 IpPort=53110 TicketEncryptionType=0x17
TicketOptions=0x40810000 Status=0x0 TransmittedServices=-Baseline the affected identity, host, application, service, or protocol.
Cloud sign-in context including identity, application, source, result, device, and conditional access.
2026-07-16T14:24:51Z UserPrincipalName=j.singh@corp.example
AppDisplayName="Office 365 Exchange Online" IPAddress=203.0.113.48
ResultType=50126 ResultDescription="Invalid username or password"
ClientAppUsed=Browser ConditionalAccessStatus=notApplied
DeviceDetail.operatingSystem=Windows Location=SGDetect rare combinations of source, target, action, timing, volume, and result.
Failed authentication; useful for guessing, spraying, exposed remote services, and account targeting.
2026-07-16T14:22:31Z EventID=4625 Computer=DC01.corp.example
TargetUserName=j.singh TargetDomainName=CORP LogonType=3
AuthenticationPackageName=NTLM WorkstationName=WKSTN-442
IpAddress=203.0.113.48 IpPort=51842
Status=0xC000006D SubStatus=0xC000006A FailureReason="Unknown user name or bad password"Correlate identity, endpoint, network, application, and control-plane evidence across the sequence.
- Who or what identity acted?
- From which device, process, IP, or workload?
- What target and operation were involved?
- Did it fail, succeed, or change state?
- What correlated event happened immediately before and after?
Detection logic / SIEM queries
Baseline the affected identity, host, application, service, or protocol.
Detect rare combinations of source, target, action, timing, volume, and result.
Correlate identity, endpoint, network, application, and control-plane evidence across the sequence.
Prioritize activity followed by successful access, privilege change, execution, or data movement.
FROM security_events
WHERE event_time > now() - 30m
AND behavior = "Golden SAML"
GROUP BY source, target, user
HAVING count(*) > baseline(source, target, user)
OR rare(action, source, target) = true
CORRELATE WITH success, privilege_change, execution, data_accessMitigations
Reduce exposed attack surface and remove unused services, identities, features, and trust relationships.
Apply least privilege, strong authentication, secure defaults, segmentation, and timely patching.
Validate untrusted input and independently authorize sensitive operations.
Centralize and protect relevant logs; test detections with controlled simulations.
Maintain a response playbook for containment, credential rotation, evidence preservation, and recovery.
Real-world examples
Golden SAML appears in opportunistic campaigns and targeted intrusions when the required condition exists.
Real incidents usually combine it with credential theft, phishing, exploitation, persistence, or exfiltration.
The best case studies describe the full attack chain rather than an isolated tool or indicator.
Common questions
01What makes Golden SAML possible?
A trust assumption fails: an identity, input, component, network message, user decision, or software relationship is accepted without enough verification.
02What should a defender collect first?
Start with logs closest to the decision point, then add identity, endpoint, network, and control-plane context.
03How should I study this attack?
Learn the concept, identify prerequisites, map observable steps, write a detection hypothesis, and validate it safely in a lab.