When Your AI Agent Hands You the Malware: Inside FakeGit's AgentBaiting Campaign

Amber | Atatck
Download Now
When Your AI Agent Hands You the Malware: Inside FakeGit's AgentBaiting Campaign | TA2026209 | HivePro Threat Advisory

Threat Advisory • Attack Report • TLP: Amber

When Your AI Agent Hands You the Malware: Inside FakeGit's AgentBaiting Campaign

FakeGit is a large-scale operation of roughly 7,600 malicious GitHub repositories that disguise SmartLoader malware as ordinary software downloads, with over 800 posing as AI Skills or MCP servers; its AgentBaiting evolution sees AI agents independently discover these repositories and relay the attacker's installation instructions to their users.

Campaign: FakeGit AgentBaiting Malware: SmartLoader, StealC TLP: Amber Admiralty Code: A1 TA2026209
First Seen
March 2026
Targeted Regions
Worldwide
Targeted Platform
Windows
Targeted Users
Developers, AI Skills / MCP users
Malware
SmartLoader, StealC
Campaign
FakeGit
Admiralty Code
A1
Date of Publication
July 23, 2026
TA Number
TA2026209

Summary

FakeGit is a large-scale operation of roughly 7,600 malicious GitHub repositories, created by about 6,600 profiles, that disguise SmartLoader malware as ordinary software downloads. More than 800 of these FakeGit repositories pose as AI Skills or MCP servers and have surfaced over 600 times across public AI registries, giving them a false sense of legitimacy.

Victims who download and run the ZIP "installer" trigger a LuaJIT loader chain that deploys SmartLoader, which then establishes persistence and delivers the StealC information stealer. The campaign's most notable evolution, dubbed AgentBaiting, sees AI agents independently discover these FakeGit repositories during routine tasks and relay the attacker's installation instructions to their users.


Attack Details

#1 A Continuation of Water Kurita's Lumma Stealer Operation

FakeGit works by making malware look like a routine software install, and it is not new. It is assessed as a continuation of an older operation that pushed Lumma Stealer, tracked by Trend Micro under the threat actor name Water Kurita. Operators copy or fabricate GitHub projects and host them under lookalike developer identities, often just one character off a real handle, then dress them up with fake star and fork counts, copied descriptions, and polished READMEs. Many impersonate familiar tools like Gmail, WhatsApp, Databricks, Jenkins, and Docker, and more than 800 pose as AI Skills or MCP servers, amplified by over 600 listings across public registries such as LobeHub, Glama, MCP.so, and MCP Market. The key evolution is AgentBaiting: an AI agent hunting for a capability can find one of these FakeGit repositories on its own, treat the attacker's README as legitimate documentation, and pass the install steps to its user. In Island's testing, Claude Code, Gemini, and ChatGPT all surfaced malicious repositories without being shown a link.

#2 A LuaJIT Loader Chain Disguised as a Software Install

The FakeGit attack starts the moment the victim opens the downloaded ZIP. A representative package holds no real software, just three files: a small launcher script (for example, application.cmd), a renamed LuaJIT-style runtime (luau.exe), and an obfuscated Lua payload disguised as a text or icon file (ico64.txt). The launcher passes the payload to the runtime, executing a roughly 300 KB single-line Lua program. Names vary across the campaign, but the structure is constant: a .cmd or .bat launcher, a LuaJIT-style runtime, and a payload masquerading as a text, icon, license, or data file. This loader chain drops SmartLoader, which hides its console window and persists through scheduled tasks under %LOCALAPPDATA%.

#3 A Blockchain Dead Drop Feeds SmartLoader's C2

Instead of moving laterally, the FakeGit campaign focuses on the endpoint and quietly stages its next payload. SmartLoader pulls its current command-and-control address from a value stored in a Polygon smart contract. This blockchain dead drop makes takedown harder and then retrieves encrypted stages from GitHub. Those stages lead to a PE crypter that injects the StealC information stealer into another process. Because the host is already compromised at execution, the fake "integration" never needs the access it advertised.

#4 StealC Harvests Live Sessions, Not Just Passwords

StealC handles the theft once it is running in memory. It targets browser passwords, cookies, active sessions, extension data, email and remote-access credentials, screenshots, and host details, then exfiltrates them over its C2 channel. Crucially for responders, StealC steals live sessions, not just stored passwords, so resetting credentials is not enough. Active browser sessions, OAuth grants, API tokens, and cloud and developer credentials must all be revoked to fully close out a FakeGit-related intrusion.


Recommendations

01

Build a Curated Capability Catalog

Maintain an approved, internally reviewed catalog of Skills, MCP servers, and agent plugins so users and agents pull capabilities from a trusted source instead of open, unvetted discovery, which is exactly the behavior the FakeGit campaign depends on.

02

Sandbox Every New Capability Before Rollout

Evaluate any new Skill or MCP server in an isolated environment that has no browser sessions, cloud credentials, SSH keys, or production data before allowing it into wider use.

03

Reject Executable-Delivered Capabilities on Sight

Treat any Skill or MCP server that installs through a downloaded executable or ZIP launcher as malicious. A legitimate capability is source code with a manifest, not a Windows ZIP holding a launcher script, a renamed runtime, and a payload disguised as a text file.

04

Verify the Publisher, Not Just the Project

Independently confirm developer identities before trusting a repository. The FakeGit campaign relies on one-character-off usernames, copied profiles, and modest star counts, so treat star and fork counts as unreliable signals.

05

Treat Registry Listings as No Signal of Trust

Do not assume a Skill or MCP server is safe because it appears on LobeHub, Glama, MCP.so, MCP Market, or similar catalogs. These registries reproduced FakeGit READMEs and download links, adding a false layer of credibility.

06

Monitor Agent-Initiated Activity

Apply the same scrutiny to downloads, git clones, and shell commands originating from AI agents as you would to user-initiated browser downloads. AgentBaiting means an agent completing a routine task can retrieve and stage SmartLoader malware on its own.

07

Maintain an AI Capability Inventory with Hashes

Keep an inventory of every AI capability in use, including source repository, commit, version, and package hash, so newly disclosed FakeGit campaign repositories can be matched against your environment quickly.


Potential MITRE ATT&CK TTPs

T1585
Resource Development
Establish Accounts
T1587.001
Resource Development
T1587: Develop Capabilities — Malware
T1608.001
Resource Development
T1608: Stage Capabilities — Upload Malware
T1195
Initial Access
Supply Chain Compromise
T1204.002
Execution
T1204: User Execution — Malicious File
T1059.003
Execution
T1059: Command and Scripting Interpreter — Windows Command Shell
T1036.005
Defense Evasion
T1036: Masquerading — Match Legitimate Name or Location
T1027
Defense Evasion
Obfuscated Files or Information
T1140
Defense Evasion
Deobfuscate/Decode Files or Information
T1564.003
Defense Evasion
T1564: Hide Artifacts — Hidden Window
T1055
Defense Evasion
Process Injection
T1053.005
Persistence
T1053: Scheduled Task/Job — Scheduled Task
T1102.001
Command and Control
T1102: Web Service — Dead Drop Resolver
T1105
Command and Control
Ingress Tool Transfer
T1555.003
Credential Access
T1555: Credentials from Password Stores — Credentials from Web Browsers
T1539
Credential Access
Steal Web Session Cookie
T1113
Collection
Screen Capture
T1005
Collection
Data from Local System
T1041
Exfiltration
Exfiltration Over C2 Channel

Indicators of Compromise (IOCs)

TypeValue
SHA256216a2c99fd42c00f9323d8b16dd19f622f7f4778b2b1d7cf07a3de5621fd1546
SHA25691e5dbfaf45edf25fbc2168f92083e05dfa427afa7633e991392e33cc7427dad
SHA256498fe8fb806cd0e6685f97fc7d74de769dae5a28cdc821557b7585ad5ad83147
SHA25662744baa8077bb8be237647fd78e3bea2ca0932bf4be3d5618600f97118095f8
SHA2561da8df487d30b988f3c350c065206726aaa13f079a07151cd42ab5579994b9de
SHA256c15693106682f2ddb26649cab6e1962a64537627cde4c5d3c79d5a0be8c1b5a8
SHA25666afc7d87d10dbe392898c4e5c613e0442fabb396415c2bef3a5ef2ac752c5ad
SHA256a33f40cab1ab7f971d3464af3e7595918107332b9e83342007571842b9e22826
SHA2563c858facbad66f5479e2c4add171421dc1b6488b36f33e7cff073aba585954a7
SHA256fc1278f419e611bf40ca414099bfd9ad98a31ffb054371e8cb65a84849b00eaf
Filenameyu-ai-agent-1.0-beta.3.zip
Filenameawesome-skills-claude-3.3.zip
Filenameagent_ai_awesome_skills_2.0.zip
Filenamefor-security-mcp-3.3.zip
Filenameplugins_claude_awesome_code_2.4.zip
Filenamemcp-walmart-2.2.zip
Filenameserver_databricks_mcp_1.6.zip
Filenamemcp-server-jenkins-3.2.zip
Filenamegateway-docker-mcp-v1.6-alpha.5.zip
Filenamealibabacloud-skills-bigdata-v1.7.zip
GitHub Repositoryhfgwygey/yu-ai-agent
GitHub RepositoryMann1988/awesome-claude-skills
GitHub Repositoryh4vzz/awesome-ai-agent-skills
GitHub RepositoryStanLeyJ03/mcp-for-security
GitHub Repositoryxbim08/awesome-claude-code-plugins
GitHub RepositoryDomingosNgongo/walmart-mcp
GitHub Repository45d5r/databricks-mcp-server
GitHub RepositoryMauManto/jenkins-mcp-server
GitHub Repositorywaynestimulative605/docker-mcp-gateway
GitHub Repositorylucaducapuca/alibabacloud-bigdata-skills
GitHub RepositoryNaveenkm007/spaceship-mcp
GitHub Repositoryhahaha-saygex/gmail-mcp
GitHub Repositoryadlaiponderous700/claude-skill-cinematic-prompt

References


What Next?

At Hive Pro, it is our mission to detect the most likely threats to your organization and to help you prevent them from happening.

Book a Demo of HivePro.