Book a demo →
← All posts
Threat Watch

SOMBERMEME: Altered Claude Code Settings Fetched a Backdoor

Reports on 16 and 17 Nov 2026 say a suspected North Korean cluster used fake recruiters and ClickFix, then changed Claude Code settings to fetch SOMBERMEME.

SOMBERMEME: Altered Claude Code Settings Fetched a Backdoor

AegisAI Threat Watch · ClickFix and developer tooling

SOMBERMEME: Altered Claude Code Settings Fetched a Backdoor

News reports published on 16 and 17 November 2026 say a suspected North Korea-linked cluster used fake recruiters and ClickFix lures against cryptocurrency organizations, then changed Claude Code settings so that a backdoor downloaded when the tool launched.

TL;DR
What
A suspected North Korea-linked cluster used fake recruiters and ClickFix, then changed Claude Code settings so a backdoor downloaded on launch.
Who it hits
Employees at cryptocurrency organizations in North America and Europe, as reported.
Why filters miss it
The fake recruiter profiles were on professional networking sites, and the victim ran the ClickFix script themselves.
Do today
Review AI coding tool settings files in repositories and ban pasting commands from web pages.

Reported by Help Net Security, BleepingComputer and The Record on 16 and 17 November 2026.

SOMBERMEME is the name given, in the reports, to a backdoor that was set to download when a victim launched Claude Code. The reports say a suspected North Korea-linked cluster used AI-generated recruiter profiles and ClickFix lures against cryptocurrency organizations in North America and Europe, then changed Claude Code's settings.

ClickFix is a technique in which a fake error or check tells the visitor to paste and run a command themselves. Claude Code is Anthropic's AI coding assistant, which developers run on their own machines. .claude/settings.json is the settings file that Help Net Security named. A GitHub Personal Access Token is a credential that lets a script or a person act as a GitHub user.

Analysis based on three news reports of a threat intelligence team's findings, published on 16 and 17 November 2026. This article does not cite the original research and rests on the three reports' wording. The reports give no victim count, no dates for the intrusions and no indicators. A related post covers a fake Zillow ClickFix chain.

2
regions where cryptocurrency organizations were targeted: North America and Europe
2
names in the reports for the cluster: MIDNIGHT NEPTUNE and UNC6836
1
settings file named: .claude/settings.json
3
news reports behind this article

Source: Help Net Security, BleepingComputer and The Record, 16 and 17 November 2026. UNC6836 is described by BleepingComputer as a cluster with suspected links to Midnight Neptune.

Key judgments

  • The attribution is hedged. Help Net Security calls the actor "a suspected cluster tracked as MIDNIGHT NEPTUNE". BleepingComputer names UNC6836, a cluster with suspected links to Midnight Neptune. The Record and BleepingComputer describe the actors as North Korean, with BleepingComputer saying "suspected".
  • The approach used a made-up person. Help Net Security reports fraudulent recruiter profiles whose photos carried SynthID watermarks, which confirmed they were AI-generated. Targets were steered to bogus meeting platforms with ClickFix lures that asked them to run "diagnostic" scripts.
  • A settings change started the backdoor. Help Net Security reports that attackers changed .claude/settings.json files in internal code repositories so that launching Claude Code triggered the SOMBERMEME download. BleepingComputer and The Record report the same launch trigger.
  • Tokens carried a trusted name. Help Net Security reports that the attackers named GitHub Personal Access Tokens they generated after Microsoft Copilot, to blend in during lateral movement. The other two reports do not mention it.
  • The detail is thin. The three reports give no victim count, no dates, no domains and no file hashes, and none describes a flaw in Claude Code itself.

Evaluating email security tools? The 2026 Buyer's Guide to Email Security has nine evaluation areas and the questions to ask vendors. Get the guide.

How did fake recruiters get cryptocurrency staff to run a script?

The reports describe a two-step approach: a recruiter profile first, then a meeting page with a ClickFix lure. The Record says North Korean actors used AI to create fake recruiter profiles on professional networking sites in recent intrusions into cryptocurrency organizations in North America and Europe, and that these personas then ran ClickFix attacks. Help Net Security adds that the profile photos carried SynthID watermarks and that targets were steered to bogus meeting platforms.

A ClickFix lure usually shows a fake problem and a command that supposedly fixes it. Help Net Security says the scripts were presented as "diagnostic" scripts. The Record notes that North Korean actors have used ClickFix widely against cryptocurrency companies, so the technique itself is not new to this group. The reports do not name the networking sites, the job on offer or what the script did when it ran.

What the reports say about the first contact

DetailReported byWhat was reported
Fake recruiter profilesThe Record, Help Net SecurityAI-generated profiles on professional networking sites; photos carried SynthID watermarks
Meeting platformHelp Net SecurityTargets were steered to bogus meeting platforms
ClickFix lureHelp Net Security, The RecordThe Record names ClickFix attacks; Help Net Security says the scripts were presented as "diagnostic" scripts
TargetsAll threeCryptocurrency organizations in North America and Europe

Source: Help Net Security (17 Nov 2026), The Record (17 Nov 2026), BleepingComputer (16 Nov 2026).

How did a changed Claude Code settings file start the SOMBERMEME download?

The reports say the attackers changed Claude Code's settings after they were already inside, so that a normal launch of the tool triggered the download. Help Net Security says the attackers moved into internal code repositories and changed .claude/settings.json files. The Record says that once ClickFix had breached a device, the actors pivoted into internal code repositories and modified code so that starting Claude Code triggered the download of a backdoor that researchers call SOMBERMEME. BleepingComputer says the attackers modified the Claude Code app's settings.

None of the three describes a flaw in Claude Code. They describe changes made after the attackers had access. The reports do not say whether the edit was made on a developer's machine, in a repository or both, and they do not say what SOMBERMEME does once it downloads. The Record also has a line about installing code into the company's language models. It is unclear and is not used here.

What each report says about the settings change

ReportDateWhat it says
Help Net Security17 Nov 2026Names .claude/settings.json; edited in internal code repositories; launching Claude Code triggered the SOMBERMEME download
BleepingComputer16 Nov 2026The attackers modified the Claude Code app's settings; the change triggered the download
The Record17 Nov 2026After a ClickFix breach, actors pivoted into internal code repositories; starting Claude Code triggered the download

Why did the attackers name GitHub tokens after Microsoft Copilot?

Help Net Security reports that the attackers named the GitHub Personal Access Tokens they generated after Microsoft Copilot, to blend in during lateral movement. It is the only one of the three reports that mentions this, and it gives no more detail about the tokens.

This reading is AegisAI's analysis. A token name is a label that whoever creates the token chooses. A reviewer scanning a list of tokens may skip a name that looks like a known assistant, and the name says nothing about who created the token or why.

Why do standard email and endpoint controls miss this chain?

This section is AegisAI's analysis. The reports describe recruiter profiles on networking sites and mention no email, and the lure sat on a meeting platform. In a ClickFix lure the victim runs the command, so the reports describe no attachment. The later steps were a file edit and a download triggered by a tool launch the user would expect. Each step looks like ordinary work.

Where each control stands

Six controls are compared against the reported chain. The reports do not test any of them.

ControlWhat it sees in this campaignResult
Email gateway and link scanningNo email stage was reportedWould only apply if a recruiter message or meeting link arrived by email
Attachment sandboxA pasted script, not an attachmentDoes not apply to the reported steps
Endpoint monitoring of pasted commandsA script run by the user in their own sessionCan help if it flags the script; the reports do not say whether any tool did
Review of changes to AI tool settings filesAn edit to .claude/settings.jsonCan expose the change if settings files are in scope of repository review
Token inventory by owner, scope and creation dateTokens named after Microsoft CopilotCan expose them when the name is checked against who created the token
Recruiter verification policyA fake recruiter profileCan stop the chain at the first step

AegisAI's reading of the reported chain. The reports do not test any of these controls.

Each step in the reported chain borrowed something ordinary: a recruiter message, a fix for a meeting problem, a tool launch and a token name.

AegisAI Threat Intelligence

What this means for a cryptocurrency or software engineering team

This section is AegisAI's analysis. Developer machines hold access to repositories and tokens, and in this campaign a settings change was enough to make a download start on launch. Exposure depends on three things: whether staff know that a recruiter should never ask for a command to be run, whether settings files for AI coding tools are reviewed like code, and whether anyone can list every GitHub token with its owner and creation date.

The reports describe cryptocurrency organizations. The same steps could be tried against any organization whose developers use a similar tool, but the reports do not say that happened.

Why standard controls miss it

This section is AegisAI's analysis. The approach starts away from the mailbox, the script is run by a person, the settings file is not an executable, and a token name can imitate a trusted tool. A control that watches only for new executables or only for email would see none of these steps.

How AegisAI protects against it

AegisAI has not observed this campaign. The reports place the fake recruiter profiles on professional networking sites and the lure on a meeting platform, not in email. The approach below applies where a recruiter message or a meeting link reaches a work mailbox.

AegisAI's agents evaluate the message and the destination it points to together, and pull a malicious message out of mailboxes before users see or click it. More on how that works: inbound email security.

When a lure or a compromised session is caught, the response compounds: the message is withheld at classification, copies already delivered are retracted, rewritten links refuse a late visit, the account's outbound and internal mail gets rescanned, and lateral phishing attempts to colleagues are blocked at the same bar rather than treated as a fresh, unrelated message.

How AegisAI stops credential phishing, 48 second video

How AegisAI stops credential phishing, 48 seconds

Recommendations for defenders

  • Set a rule that recruiters and hiring contacts are verified through the company's own channels, and that no interview step needs a command, a script or a "fix" on a meeting page.
  • Train developers on ClickFix: no meeting tool, browser error or software fix needs a command pasted into a terminal or Run dialog. Give staff one place to report such a request.
  • Treat settings files for AI coding tools, such as .claude/settings.json, as code. Review changes to them in repositories, alert on edits, and check the file on developer machines.
  • Inventory GitHub Personal Access Tokens by owner, scope and creation date, set expiry, and review any token whose name claims to belong to a known tool.
  • If a command was pasted from a page, treat the device as compromised: isolate it, rotate the credentials it held, and review repositories for settings changes and new tokens.

FAQ

What is the SOMBERMEME backdoor?

SOMBERMEME is the name used in the reports for a backdoor that downloaded when a victim launched Claude Code after attackers changed its settings. The reports do not say what the backdoor does after it downloads.

Who is MIDNIGHT NEPTUNE (UNC6836)?

Help Net Security calls the actor a suspected cluster tracked as MIDNIGHT NEPTUNE. BleepingComputer names UNC6836, a cluster with suspected links to Midnight Neptune. The Record and BleepingComputer link the activity to North Korea, and BleepingComputer says "suspected".

Is this a vulnerability in Claude Code?

None of the three reports describes a flaw in Claude Code. They describe settings changes made after the attackers already had access to a device and to internal repositories.

What should I do if a recruiter asks me to run a command?

Stop and do not run it. Tell your security team, and check the recruiter through the company's own site or a known contact. If a command was already run, tell the team at once so the device can be isolated and credentials changed. These steps are AegisAI's general guidance, not taken from the reports.

For researchers

What the reporting does not tell us 5 gaps

Scale and timing. The reports give no victim count, no intrusion dates and no list of organizations.

Attribution. The cluster name and the country link vary by report, and the Midnight Neptune link to UNC6836 is described as suspected.

Indicators. None of the three publishes meeting platform domains, SOMBERMEME file hashes or the token names used.

The settings change. The reports do not say where the edit was made, how the attackers reached the repositories beyond the ClickFix step, or what SOMBERMEME does after it downloads.

Source scope. The three reports cover a wider research report on attackers' use of AI. This article rests on their wording, and the Copilot token detail comes from one outlet. AegisAI did not test or observe any part of the chain.

Reported identifiers 4 items

The reports publish names, not network indicators. Treat these as labels for searching, not a blocklist.

TypeIdentifierRole
Malware nameSOMBERMEMEBackdoor downloaded when Claude Code launched after the settings change
Cluster nameUNC6836 / MIDNIGHT NEPTUNENames used in the reports; the link between them is described as suspected
File.claude/settings.jsonSettings file Help Net Security says was edited
Credential labelToken named after Microsoft CopilotGitHub Personal Access Tokens the attackers generated
MITRE ATT&CK mapping 5 techniques

These techniques are AegisAI's mapping of the reported behaviour. The reports do not map the campaign to technique IDs. IDs and names were checked against attack.mitre.org on 18 November 2026.

IDNameUse in this campaign
T1585.001Establish Accounts: Social Media AccountsFake recruiter profiles with AI-generated photos were set up on professional networking sites.
T1566.003Phishing: Spearphishing via ServiceFake recruiters on a networking platform steered targets to a bogus meeting platform.
T1204.004User Execution: Malicious Copy and PasteClickFix lures asked victims to run "diagnostic" scripts themselves.
T1105Ingress Tool TransferLaunching Claude Code after the settings change triggered a download of SOMBERMEME.
T1036.005Masquerading: Match Legitimate Resource Name or LocationGitHub tokens were named after Microsoft Copilot to blend in.
References 4 sources

Phishing that reaches the inbox, stopped before anyone clicks.

Book a demo