Skip to navigation

Threat reportMalwareTL-2026-0621

mouse5212-super-formatter — AI-Generated Malicious npm Package Exfiltrating Anthropic Claude AI /mnt/user-data Sandbox to Attacker GitHub Repository (Malware-Slop Campaign)

highRESOLVED

OX Security uncovered a malicious npm package, mouse5212-super-formatter, that masquerades as an internal 'archive deployment sync' utility but recursively exfiltrates the contents of Anthropic Claude AI's /mnt/user-data sandbox to a threat actor-controlled GitHub repository via the GitHub Contents API. The package executed during the npm postinstall lifecycle hook, authenticated to GitHub using either a victim-environment GITHUB_TOKEN or a hardcoded fallback Personal Access Token, base64-encoded each file, and uploaded it into a per-session random folder.

Key facts

IdentifiersTL-2026-0621
Also known asMalware-Slop, mouse5212-super-formatter Claude AI stealer
SeverityHIGH
Affected productsnpm registry (mouse5212-super-formatter maintainer) mouse5212-super-formatter (npm package); Anthropic Claude AI /mnt/user-data sandbox directory
Weaknesses (CWE)CWE-506, CWE-829, CWE-798, CWE-200
MITRE ATT&CK techniquesT1005 Data from Local System, T1027 Obfuscated Files or Information, T1036 Masquerading, T1041 Exfiltration Over C2 Channel, T1059 Command and Scripting Interpreter, T1071 Application Layer Protocol, T1083 File and Directory Discovery, T1102 Web Service, T1119 Automated Collection, T1195 Supply Chain Compromise, T1204 User Execution, T1546 Event Triggered Execution and 8 more in the attack flow
MotivationFINANCIAL
Target sectorstechnology, software-development, artificial-intelligence, research, developer-tooling
Target regionsGlobal
Detection rules9 (1 shown free)
Sources8 cited sources

Detection coverage

As of 2026-05-28, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-0621 across Splunk SPL, Microsoft KQL and Sigma, covering 21 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Burst of PUT requests to api.github.com Contents API from npm-installed node process — C2/Exfiltration Phase

Detects the exfiltration stage of the Malware-Slop campaign: a node/npm process tree generates a burst of HTTPS requests to api.github.com targeting the /repos/{owner}/{repo}/contents/ Contents API endpoint, which the attacker uses both as C2 (repo provisioning, token validation) and as the data sink. Maps to T1071.001, T1567.001, T1041, and T1102.002. Targets proxy/NGFW logs or Sysmon EventCode 3 + DNS correlation.

Splunk SPL

index=proxy sourcetype IN ("bro:http:json", "zeek:http", "cisco:asa", "pan:traffic", "nginx:proxy", "squid:access") (dest_host="api.github.com" OR url="*api.github.com*")
| rex field=url "https?://api\.github\.com/repos/(?<gh_owner>[^/]+)/(?<gh_repo>[^/]+)/contents/(?<gh_path>[^?\s]+)"
| where method="PUT" AND isnotnull(gh_owner)
| join type=left process_guid [ search index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=3 dest_host="api.github.com" (Image="*\\node.exe" OR Image="*/node") | fields process_guid Image ParentImage host User ]
| where match(ParentImage, "npm|yarn|pnpm|npx") OR match(Image, "node")
| stats count dc(gh_path) as paths_touched values(gh_owner) as repo_owners values(gh_repo) as repos min(_time) as first_seen max(_time) as last_seen by host User Image ParentImage gh_owner gh_repo
| where count >= 3
| eval threat_id="TL-2026-0621", detection="GitHub Contents API exfiltration from npm-spawned node"
| table _time host User threat_id detection ParentImage Image gh_owner gh_repo count paths_touched first_seen last_seen

Microsoft KQL

let window = 5m;
let exfil_threshold = 3;
DeviceNetworkEvents
| where Timestamp > ago(24h)
| where RemoteUrl has "api.github.com" and RemoteUrl has "/contents/"
| where InitiatingProcessFileName in~ ("node", "node.exe")
| where InitiatingProcessParentFileName in~ ("npm", "npm.exe", "yarn", "yarn.exe", "pnpm", "pnpm.exe", "npx", "npx.exe", "node", "node.exe")
| extend GhOwner = extract(@"api\.github\.com/repos/([^/]+)/", 1, RemoteUrl)
| extend GhRepo = extract(@"api\.github\.com/repos/[^/]+/([^/]+)/contents", 1, RemoteUrl)
| extend GhPath = extract(@"/contents/([^?\s]+)", 1, RemoteUrl)
| summarize Hits = count(), DistinctPaths = dcount(GhPath), Owners = make_set(GhOwner, 5), Repos = make_set(GhRepo, 5), FirstSeen = min(Timestamp), LastSeen = max(Timestamp) by DeviceName, InitiatingProcessId, InitiatingProcessParentFileName, bin(Timestamp, window)
| where Hits >= exfil_threshold or DistinctPaths >= exfil_threshold
| project FirstSeen, LastSeen, DeviceName, InitiatingProcessParentFileName, Hits, DistinctPaths, Owners, Repos
| order by FirstSeen desc

Sigma

title: Burst of GitHub Contents API PUT from npm-spawned node process
id: c8a3f721-0621-4003-ac03-mslop003
status: experimental
description: Detects the exfiltration stage of mouse5212-super-formatter — a node child of npm/yarn/pnpm emits multiple HTTPS PUT requests to api.github.com/repos/{owner}/{repo}/contents/ within a short window. This is the dual-use C2 + exfil sink of the Malware-Slop campaign.
references:
    - https://www.ox.security/blog/malware-slop-new-malicious-npm-package-leaks-its-own-github-private-token/
    - https://attack.mitre.org/techniques/T1567/001/
    - https://attack.mitre.org/techniques/T1071/001/
    - https://attack.mitre.org/techniques/T1102/002/
author: Threadlinqs Team
date: 2026/05/28
tags:
    - attack.command_and_control
    - attack.t1071.001
    - attack.t1102.002
    - attack.exfiltration
    - attack.t1567.001
    - attack.t1041
logsource:
    category: network_connection
    product: linux
detection:
    selection_proc:
        Image|endswith:
            - '/node'
        ParentImage|endswith:
            - '/npm'
            - '/yarn'
            - '/pnpm'
            - '/npx'
            - '/node'
    selection_dest:
        DestinationHostname: 'api.github.com'
    selection_uri:
        DestinationUrl|contains: '/contents/'
        HttpMethod: 'PUT'
    timeframe: 5m
    condition: all of selection_* | count() by Image > 3
falsepositives:
    - Developer tooling that legitimately writes content to GitHub repos via the Contents API (uncommon outside CI bots)
    - GitHub Actions runners executing official content-publishing actions
level: critical

Known false positives:

  • Legitimate developer/CI tools publishing content to GitHub via Contents API (e.g. release-drafter, repo-sync bots)
  • GitHub Actions runners with content-publishing workflows

Tuning notes: Allow-list known internal automation accounts and CI runner IPs. To reduce noise, gate on the repo owner being NOT in your organization's known list. Increase fidelity by joining with a list of recently-installed npm packages (lockfile diff) — alerts that line up with a fresh install are near-zero false-positive.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Malicious npm package install — [indicator] known-bad name or postinstall manifest written — Exploitation Phase · unlock with TL Creds
  • Hardcoded GitHub PAT (ghp_*) or env GITHUB_TOKEN access by npm-spawned node process — Post-Exploitation Phase · unlock with TL Creds
  • Decoy network_connections.log committed alongside base64 blobs to fresh GitHub repo — C2/Exfiltration Phase · unlock with TL Creds
  • npm/yarn/pnpm install spawns node executing in-tree postinstall.js — Exploitation Phase (Sigma) · unlock with TL Creds
  • node child of npm reads 5+ files under [indicator] in short window — Post-Exploitation Phase (Sigma) · unlock with TL Creds
  • Burst of PUT requests to [indicator] /contents/ from npm-spawned node — C2/Exfiltration Phase (Sigma) · unlock with TL Creds
  • Malware-Slop npm postinstall script triggers node child during install — Exploitation Phase · unlock with TL Creds
  • Recursive read of Claude [indicator] by npm/node lifecycle process — Post-Exploitation Phase · unlock with TL Creds

Indicators of compromise

As of 2026-05-28, Threadlinqs Intelligence tracks 21 indicators of compromise for TL-2026-0621.

Sample indicators (2 of 21)

  • Domain: registry[.]npmjs[.]org
  • Domain: api[.]github[.]com

Defanged, TLP:CLEAR. Full set of 21 indicators: unlock with TL Creds.

21 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

The hardcoded fallback token leaked the attacker's GitHub identity, allowing OX to enumerate approximately seven exfiltration events before the account was deleted. The package reached 676 downloads across versions 1.0.0 through 1.0.4 (published 2026-05-26, unpublished 2026-05-27) and is assessed by OX as an AI-generated 'malware-slop' campaign — likely LLM-authored code with sloppy operational security, a pattern researchers warn will scale.

## Overview

mouse5212-super-formatter is a malicious npm package discovered by OX Security researchers Moshe Siman Tov Bustan and Nir Zadok and disclosed on 2026-05-27 under the codename 'Malware-Slop.' The package is purpose-built to harvest files from Anthropic Claude AI's `/mnt/user-data` sandbox — the directory Claude uses to handle user uploads, downloads, and code/data outputs — and exfiltrate them to a GitHub repository owned by the attacker. Five versions (1.0.0 through 1.0.4) were uploaded to the public npm registry on 2026-05-26 and aggregated 676 downloads before the maintainer unpublished the package on 2026-05-27. The associated GitHub account was created on 2026-05-26, only hours before the first npm publish, and has since been deleted.

## Targeting

The malware specifically targets `/mnt/user-data`, the sandbox path Anthropic mounts inside Claude's code-execution environment for file uploads and code/data outputs. Files written there typically include user-supplied source code, datasets, credentials accidentally pasted into prompts, intermediate analysis artifacts, and the contents of files Claude has been asked to operate on. Any developer who installed mouse5212-super-formatter inside a Claude sandbox (for example, while having Claude install npm dependencies in a coding workflow) would have had the full sandbox state exfiltrated. The Register's reporting recommends that any user who installed the package immediately revoke GitHub access tokens and treat all `/mnt/user-data` contents as compromised.

## Exploit Chain

1. **Initial Access — Supply Chain Compromise.** The attacker registers an npm account and publishes mouse5212-super-formatter (versions 1.0.0–1.0.4) on 2026-05-26. The package presents itself as an internal 'archive deployment sync' utility. 2. **Execution — postinstall lifecycle hook.** When a victim runs `npm install`, the package.json `scripts.postinstall` field executes the malicious JavaScript automatically, without further user action. 3. **Authentication — environment token or hardcoded fallback PAT.** The script first attempts to read a GitHub token from the victim's environment variables; if absent, it falls back to a hardcoded GitHub Personal Access Token belonging to the attacker. This fallback is the OPSEC failure that enabled OX to attribute the campaign. 4. **Resource Development — repository provisioning.** Using the GitHub REST API, the script checks whether the attacker's target repository exists. If not, it creates the repository on the fly. 5. **Discovery and Collection — recursive walk of /mnt/user-data.** The script enumerates the Claude sandbox directory recursively, reading every file it can access. 6. **Defense Evasion — base64 encoding and random folder names.** Each file is base64-encoded (required by the GitHub Contents API for binary content). Files are uploaded into a per-execution random folder name so that multiple stealing sessions remain distinguishable to the operator. Comments and commit messages are 'intentionally bland' English to avoid the telltale verbosity or Russian-language artifacts that often expose LLM-generated malware. 7. **Command and Control / Exfiltration — GitHub Contents API.** Files are uploaded with HTTP PUT requests to `https://api.github.com/repos/{owner}/{repo}/contents/{path}`. This dual-uses the GitHub web service for both C2 (token validation, repo existence checks, repo creation) and exfiltration (file uploads). 8. **Cover Story — fake network status log.** Alongside the stolen files, the script writes a fake 'network connections' log to the repo, intended to make the activity look like benign diagnostic telemetry rather than data theft.

## OPSEC Failure and Attribution

The hardcoded fallback PAT was embedded in plaintext in the package's JavaScript. OX Security extracted the token, used GitHub API endpoints to identify the owning account, observed approximately seven distinct exfiltration sessions (each represented as a random folder under the same destination repo), and published the findings. The attacker's GitHub account had been registered on 2026-05-26 (the same day as the npm publish), and OX additionally notes the operator first tested the stealer logic against a 'test' repository before going live — additional evidence of a hurried, AI-assisted workflow rather than a polished tradecraft.

## AI-Generated Malware Signal

OX Security attributes the code to LLM authorship based on several indicators: structurally clean function decomposition with no human-style idiosyncrasies, consistent JSDoc-style commentary, the choice of a long English package name with a stylistically odd numerical infix ('mouse5212'), and the failure to redact secrets — a class of mistake characteristic of an operator copy-pasting model output without review. OX explicitly frames the campaign as the leading edge of an emerging 'malware-slop' trend: low-effort, LLM-generated supply-chain malware that will proliferate until npm-side automated detection catches up. The Register's coverage echoes the framing and notes that the malware avoids the usual AI tells (redundant comments, Russian-language artifacts) but still leaked its operator identity in the most basic way possible.

## Affected Population

npm reports 676 download events across the five versions. OX cautions that not all downloads correspond to actual installs (mirrors, scanners, and CI dry-runs inflate the count), but every successful `npm install` of an affected version inside an environment with `/mnt/user-data` accessible would have triggered exfiltration. The package has been unpublished from npm, but developers should audit lockfiles for any of the five malicious versions and rotate any tokens that may have been present in the affected sandbox.

MITRE ATT&CK techniques by tactic:

Collection

T1005 Data from Local System; T1119 Automated Collection; T1560 Archive Collected Data

Defense Evasion

T1027 Obfuscated Files or Information; T1036 Masquerading

Exfiltration

T1041 Exfiltration Over C2 Channel; T1567 Exfiltration Over Web Service

Execution

T1059 Command and Scripting Interpreter; T1204 User Execution

Command and Control

T1071 Application Layer Protocol; T1102 Web Service; T1573 Encrypted Channel

Discovery

T1083 File and Directory Discovery

Initial Access

T1195 Supply Chain Compromise

Persistence

T1546 Event Triggered Execution

Credential Access

T1552 Unsecured Credentials

Resource Development

T1583 Acquire Infrastructure; T1585 Establish Accounts; T1608 Stage Capabilities

defense-impairment

T1685 Disable or Modify Tools

Remediation

Patches

  • npm has unpublished all versions of mouse5212-super-formatter (1.0.0 through 1.0.4) as of 2026-05-27 — no install path remains via the registry
  • Anthropic Claude AI sandboxes should treat any package install from the public npm registry as untrusted until npm rolls out automated malware blocking

Immediate actions

  • Audit package-lock.json, yarn.lock, and pnpm-lock.yaml across all repositories for mouse5212-super-formatter at any version (1.0.0, 1.0.1, 1.0.2, 1.0.3, 1.0.4)
  • If installed, treat the affected environment as compromised: revoke every GitHub Personal Access Token, fine-grained PAT, OAuth app token, and GitHub App installation token reachable from that environment
  • Treat all contents of /mnt/user-data (or the local equivalent) at time of install as exfiltrated; rotate any secrets, API keys, or credentials that may have been present
  • Remove the package from any caches: npm cache clean --force, and purge offline mirrors and Verdaccio caches
  • Review GitHub audit logs for unexpected repository creations and Contents API PUT activity attributed to revoked tokens

Workarounds

  • Run npm install with --ignore-scripts inside any Claude sandbox or comparable agentic-AI execution environment
  • Pre-scan dependencies with Socket/Snyk before allowing install in a sandbox with sensitive file mounts
  • If immediate install is required, use npm pack and inspect package contents (especially package.json scripts) before execution

Longer-term hardening

  • Block postinstall, install, and preinstall lifecycle scripts in CI by default using --ignore-scripts (npm), enableScripts false (yarn), or ignore-scripts=true (pnpm)
  • Adopt allow-listed dependency installation in sensitive AI sandboxes — pin and pre-vet every npm package before it can run in a Claude /mnt/user-data context or comparable environment
  • Egress-restrict AI code-execution sandboxes so api.github.com is reachable only when explicitly required by the user task, and log every outbound HTTPS request
  • Deploy npm supply-chain monitoring (Socket, Snyk Advisor, Phylum, OX Security) with alerting for newly registered packages installed by internal developers
  • Run secret-scanning hooks (gitleaks, trufflehog) against installed node_modules contents to catch embedded tokens before postinstall execution

Affected and fixed versions

  • npm registry (mouse5212-super-formatter maintainer) — mouse5212-super-formatter (npm package)
    Vulnerable versions: 1.0.0; 1.0.1; 1.0.2; 1.0.3; 1.0.4
    Fixed in: unpublished from npm 2026-05-27
  • Anthropic — Claude AI /mnt/user-data sandbox directory
    Vulnerable versions: any Claude code-execution session that installed the malicious package
    Fixed in: package removed from npm; sandbox unaffected when --ignore-scripts is enforced

Timeline

  • Approximately seven distinct exfiltration sessions land in the attacker repository, each under a randomly named folder; OX traces them via the leaked hardcoded GitHub PAT.
  • Versions 1.0.1, 1.0.2, 1.0.3, and 1.0.4 published within a four-hour window (18:04–18:17 UTC), consistent with hurried iteration on the postinstall payload.
  • mouse5212-super-formatter version 1.0.0 published to the public npm registry at 17:30:57 UTC.
  • OX Security observes the operator first testing the information-stealing logic against a 'test' GitHub repository in the same attacker account.
  • Attacker registers GitHub account that will host exfiltrated files; account is created only hours before the first malicious npm publish.
  • Attacker-controlled GitHub account is deleted, severing access to evidence of stolen files; revoked PAT no longer accepted by GitHub API.
  • mouse5212-super-formatter unpublished from npm at 17:53:41 UTC; total download count at removal is 676 across all five versions.
  • OX Security publishes Malware-Slop research; The Hacker News and The Register pick up the story the same day.
  • As of 2026-05-29, this specific Malware-Slop campaign is neutralized: npm registry metadata confirms all five mouse5212-super-formatter versions were unpublished 2026-05-27, and the throwaway attacker GitHub account is deleted with its hardcoded PAT revoked, killing the exfil sink. The actor is unknown/single-use with no rebrand or successor package reported, though OX warns the AI-generated postinstall-stealer technique class will spawn copycats.

Sources cited

Previous and next in malware

Further reading

Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats