MITRE ATT&CK Enterprise · TA0006
Credential Access: how an intruder takes the keys
Credential AccessTA0006
A password is an asset with a price. This is the tactic that turns one stolen password into every password the intruder can reach — what it costs, what it looks like from the outside, and the telemetry that catches it.
For the owner
An intruder who wants one thing does not need a breach. One password, one saved session, one reusable code on a screen, and the work of months is available in a minute.
That is what makes credentials the most valuable thing in a small or mid-sized company, and it is not a close comparison. A password is not a piece of information; it is a key that opens the door to a payroll system, a customer database, a shared drive of contracts, and — usually — the email account of the person who owns it, which is the account everyone else can be reset from.
The cost of a stolen credential is also the part most organisations get wrong. They treat the incident as the moment the password was taken, and they rotate it, and they consider the matter closed. The damage is not the theft. The damage is everywhere that password was already used, on every system that trusted it, for as long as it was trusted. Resetting the password does not undo a week of access.
Three questions are worth asking your own team:
- Which accounts would hurt us most if they were used by someone else tomorrow? Not the administrator accounts everybody worries about — the quiet ones: a shared mailbox, a service account, a payroll login, an account with no multi-factor on it because it was too much trouble to set up.
- Where do our people and our software keep passwords today? Spreadsheets, shared documents, chat messages, a notes application, a password written on a whiteboard in the server room. Every one of those is a credential store that somebody else can reach.
- If one credential was taken, how quickly could we find every place it was used afterwards — and how quickly could we revoke it? If the honest answer is "we would not know where to look", that is the gap that turns one stolen key into a breach you have to disclose.
None of this is about asking staff to memorise better passwords. It is about knowing which accounts matter, where the secrets live, and whether you would see the first use of something you did not authorise.
For the analyst
The tactic
MITRE ATT&CK files Credential Access under TA0006. The adversary's objective is material that authenticates as somebody: a password, a hash, a token, a certificate, a session, or a captured keystroke. The matrix files 68 techniques under it, and the pinned SigmaHQ corpus carries 323 rules that name techniques in this tactic.
Two properties make this tactic different from the others in the series. First, it is not about the machine — a credential taken from a laptop can unlock a server the intruder has never touched. Second, the theft is almost always invisible at the moment it happens and obvious afterwards, which means the durable detection is not the theft. It is the use.
The techniques the corpus detects
| Technique | Name | Rules in the corpus |
|---|---|---|
T1003.001 | OS Credential Dumping: LSASS Memory | 73 |
T1003 | OS Credential Dumping | 34 |
T1003.002 | OS Credential Dumping: Security Account Manager | 28 |
T1003.003 | OS Credential Dumping: NTDS | 23 |
T1110 | Brute Force | 23 |
T1552.001 | Unsecured Credentials: Credentials In Files | 20 |
T1558.003 | Steal or Forge Kerberos Tickets: Kerberoasting | 16 |
T1528 | Steal Application Access Token | 14 |
T1003.004 | OS Credential Dumping: LSA Secrets | 12 |
T1552 | Unsecured Credentials | 12 |
T1556 | Modify Authentication Process | 12 |
T1557.001 | Adversary-in-the-Middle: Name Resolution Poisoning and SMB Relay | 10 |
T1040 | Network Sniffing | 9 |
T1557 | Adversary-in-the-Middle | 9 |
T1003.005 | OS Credential Dumping: Cached Domain Credentials | 8 |
T1555 | Credentials from Password Stores | 8 |
T1003.006 | OS Credential Dumping: DCSync | 7 |
T1187 | Forced Authentication | 7 |
T1555.003 | Credentials from Password Stores: Credentials from Web Browsers | 7 |
T1552.004 | Unsecured Credentials: Private Keys | 6 |
T1552.006 | Unsecured Credentials: Group Policy Preferences | 5 |
T1558 | Steal or Forge Kerberos Tickets | 5 |
T1110.001 | Brute Force: Password Guessing | 4 |
T1212 | Exploitation for Credential Access | 4 |
T1552.002 | Unsecured Credentials: Credentials in Registry | 4 |
T1552.007 | Unsecured Credentials: Container API | 4 |
T1555.004 | Credentials from Password Stores: Windows Credential Manager | 4 |
T1649 | Steal or Forge Authentication Certificates | 4 |
T1056.001 | Input Capture: Keylogging | 3 |
T1056.002 | Input Capture: GUI Input Capture | 3 |
T1552.003 | Unsecured Credentials: Shell History | 3 |
T1556.002 | Modify Authentication Process: Password Filter DLL | 3 |
T1556.006 | Modify Authentication Process: Multi-Factor Authentication | 3 |
T1056 | Input Capture | 2 |
T1539 | Steal Web Session Cookie | 2 |
T1621 | Multi-Factor Authentication Request Generation | 2 |
T1110.002 | Brute Force: Password Cracking | 1 |
T1555.001 | Credentials from Password Stores: Keychain | 1 |
T1555.005 | Credentials from Password Stores: Password Managers | 1 |
T1556.004 | Modify Authentication Process: Network Device Authentication | 1 |
T1557.003 | Adversary-in-the-Middle: DHCP Spoofing | 1 |
T1606 | Forge Web Credentials | 1 |
Four ways credentials come out
Dumping them out of memory and stores (T1003, T1003.001, T1003.002, T1003.003, T1003.004, T1003.005, T1003.006). Reading the authentication service memory (T1003.001), the security account manager (T1003.002), the LSA secrets store (T1003.004), the directory database (T1003.003), cached domain credentials (T1003.005), or asking a domain controller to replicate them (T1003.006). This is the largest block of rules in the tactic, and the reason is worth understanding: the artefact is observable. A process reading the credential store produces a process-access event, and a directory replication request from a machine that is not a domain controller produces a network event with a source and destination. The normalised columns that carry this are process_name, parent_process, command_line and user_name for the local half, and src_ip, dst_ip and dst_port for the replication half. A memory dump written to a file is also a file-creation event, which is why the combination of a dump tool and a new file in a temporary directory is such a reliable pair.
Taking them where they are stored (T1555, T1555.001, T1555.003, T1555.004, T1555.005, T1552, T1552.001, T1552.002, T1552.003, T1552.004, T1552.006, T1552.007). Browser password stores, the macOS keychain, the Windows credential manager, password managers, private keys left in files, credentials in the registry, credentials in shell history, group policy preferences that still contain old passwords, and container API tokens. These are unexciting techniques with an unexciting detection problem: the read looks like a read. The signal is in the object — a file_path under a browser profile or a key directory, a process_name opening a credential store, a command_line referencing a key file.
Forcing them out (T1187, T1110, T1110.001, T1110.002, T1212, T1557, T1557.001, T1557.003, T1040). Forced authentication sends a request to a host the attacker controls so the captured response can be relayed; brute force guesses and cracks; exploitation for credential access targets the weakness directly; name resolution poisoning and SMB relay and DHCP spoofing position the attacker between two hosts that trust each other; sniffing reads the traffic. The telemetry here is the identity and network streams rather than the endpoint: repeated authentication failures for one account from one address in keycloak or m365_audit, an outbound authentication attempt to an unusual destination in firewall_syslog, a src_ip that has never initiated this conversation before.
Stealing the session instead of the secret (T1528, T1539, T1606, T1056, T1056.001, T1056.002, T1621, T1649, T1558, T1558.003, T1556, T1556.002, T1556.006). Application access tokens, web session cookies, forged web credentials, keylogging and graphical input capture, multi-factor request generation, stolen authentication certificates, and Kerberos ticket theft or forgery. Password filter libraries and multi-factor changes appear here as well, because changing the authentication process is how an intruder makes the next capture easier. The important property of this family is that the second factor is bypassed without ever being challenged: a stolen session cookie is used without a prompt, a stolen token is used without a prompt, and a stolen Kerberos ticket is used without a prompt.
Where the telemetry comes from
windows_sysmon— process access events (who opened what), process creation with the full command line, registry access, and file creation. This is the source for the dump and store families. Fields:process_name,parent_process,command_line,file_path,user_name.linux_auditd— execution records with argument vectors, privilege transitions, and access to the account database files.keycloak,m365_audit,gworkspace— failed and successful sign-ins, token grants, password resets, multi-factor changes, session activity. For this tactic this is often the most valuable stream, because it is where the use of a credential is visible.firewall_syslog— repeated failures against an external destination, and outbound authentication attempts to addresses the host has never talked to.cloud— key creation, role assumption and API token issuance.
What the numbers do and do not tell you
323 rules, and the distribution explains itself: the largest block is memory dumping, because the artefacts are visible in process and network telemetry and because the technique names the tool the attacker used. The brute-force family is small in rules and enormous in business consequence, because a rule can tell you that someone is guessing and cannot tell you which passwords will fall. That is a gap rules do not close.
Coverage is also biased by what is observable. Rules for dumping describe the well-known artefacts, not novel variants, and a variant that reads the same store through a different path may match nothing. This is the argument for detecting the access and the use rather than the tool name.
Detection engineering notes
- Anchor on the access, not the tool. "A process read the credential store" is the signal; "a known dump tool ran" is a signature that ages badly. The normalised columns make the first one easy:
file_pathanduser_nameon the accessing process. - Correlate the theft with the use. A dump is only interesting when the extracted credential is used — from a new address, on a new host, over a new protocol. A detection that fires on the theft alone produces a busy alert stream and a weak story.
- Weigh the identity stream above the endpoint. Most credential abuse is visible where the credential is used, not where it was taken.
- Watch the first use. Most of these rules describe techniques the attacker already knows. The durable detection is the behaviour that follows: a session from an unfamiliar address, an account used at an unusual hour, a service account that suddenly authenticates like a person.
- Tune for service accounts and agents. Backup software, monitoring agents, deployment tools and scheduled jobs legitimately read credential stores. Without baselines, the dump family is the noisiest in the corpus.
- Do not treat multi-factor as the end of the story. It protects the login. It does not protect the session, the token, the ticket or the certificate, and an intruder who takes one of those does not need it.
Where analysts get this wrong, and what a false negative looks like
The most common error is equating multi-factor authentication with credential protection: it stops a stolen password at the door, and does nothing about a session already open. The second common error is watching the endpoint closely and the identity provider hardly at all, in a tactic where the interesting event usually happens in the second system. The third is assuming the rule set catches novel credential tools, when it catches the well-known artefacts and nothing else.
The false negative to hold in mind is the quiet one. An intruder extracts a session cookie from a browser on a workstation, and uses it from a new address. No password was guessed, so there are no failed sign-ins. No multi-factor prompt was shown, so there is no denied challenge. No credential was wrong, so no rule about wrong credentials fires. The endpoint records a normal request. The only signal in the entire system is an existing session arriving from an address that has never used it — which is detectable, but only if the identity stream carries that context and only if somebody is looking for it.
Where DetectMonkey fits
The hard part of a credential incident is not the theft; it is the seventy-two hours afterwards. A password taken on Monday may sit in a mail archive on Tuesday, in a backup on Wednesday and in a customer's support ticket on Thursday, and rotating it on Monday evening changes none of that. What reduces the damage is knowing where the credential went and revoking it there — which is a search problem, and search problems are answered by keeping the history queryable rather than by hoping the first alert was complete.
DetectMonkey answers it that way. Endpoint, identity and firewall telemetry arrive in one normalised stream, the pinned corpus of 3164 detection rules is evaluated as the events arrive, and the last 72 hours of telemetry stay queryable behind every alert, so an analyst can ask where a credential was used and act before the window closes. Alerts carry their evidence in English and French. The platform detects and alerts. It does not isolate, quarantine, block or remediate anything on your systems, and it does not change your configuration. Telemetry is processed and stored on private servers operated by Codein in New Brunswick, Canada — not in a public cloud.
Early access is 90 days at no charge, for up to 100 seats, with founder-led onboarding and no card required.
See it on your own telemetry
Early access is 90 days at no charge, for up to 100 seats, founder-led onboarding, no card required.