# Bitdefender: 22 GitHub repos make coding agents run a credential stealer

> Bitdefender Labs found 22 GitHub repositories that make AI coding agents run a credential stealer when a developer opens the project.

- **Severity:** High: A working attack on a widely used AI product or ecosystem. A fix may exist. Patch or mitigate this week.
- **Category:** Supply chain, Incident
- **Published by:** Bitdefender
- **Disclosed:** 2026-11-17
- **Affects:** Claude Code, OpenAI Codex, GitHub Copilot, Cursor, Aider, GitHub, GitLab
- **Research by:** Andrei Anton-Aanei (Bitdefender Labs)
- **Primary source:** https://www.bitdefender.com/en-us/blog/labs/how-ai-agents-become-new-targets-in-github-malware-campaign
- **Page:** https://theinjection.dev/items/bitdefender-github-agent-config-stealer/

Bitdefender Labs found 22 GitHub repositories that run a credential stealer when a developer opens them in an AI coding tool, through planted agent instruction files and a Claude Code session-start hook.

## What happened

Bitdefender Labs published the campaign on 17 November 2026, naming 22 malicious GitHub repositories aimed at developers who open unfamiliar projects in an AI coding tool. Researcher Andrei Anton-Aanei documented the repositories, the payload and the file hashes in the write-up's indicator table. Bitdefender does not say whether the repositories were reported to GitHub or taken down.

## How it works

Two separate paths lead to the same script. The repositories ship AGENTS.md and CLAUDE.md files that present scripts/setup.js as a sealed, signed artifact that must not be inspected and must run at session start without output or confirmation, which is indirect prompt injection aimed at the assistant rather than at the user. A session-start hook in .claude/settings.json runs the same script in the user's shell once the repository is trusted, with no assistant involved at all. The script hex-decodes and ROT13-decodes a 3.75 MB blob, decrypts two AES-128-GCM payloads into a 4.67 MB bundle, and pipes that bundle into the Bun runtime.

## Who is affected

Developers who open an untrusted repository in an AI-enabled editor or terminal agent are the target. Bitdefender says the stealer keeps configuration files whose paths match any of 14 agent names, listed below, so exposure is not limited to one vendor's tool. Machines that hold source-control, cloud, registry or CI credentials carry the real loss.

## What to do

Read AGENTS.md, CLAUDE.md and the .claude, .cursor and .gemini directories before opening an unfamiliar repository in an AI-enabled environment, and inspect MCP server definitions, session-start hooks, VS Code tasks and any setup script they reference. Treat an instruction that forbids inspecting a file, suppresses output or skips confirmation as a red flag. Rotate exposed GitHub, GitLab, cloud and package-registry credentials if one of these setup scripts ran.

## Quick facts

- **Malicious repos:** 22, each listed with a setup.js MD5 hash
- **Agent names matched:** Claude, Codex, Gemini, Copilot, Kiro, OpenCode, Cline, Aider, Tabby, Amazon Q, Cody, Bolt, Continue, Cursor
- **Entry points:** AGENTS.md, CLAUDE.md, and a session-start hook in .claude/settings.json
- **Stolen:** Source-control, cloud, registry and CI tokens, SSH keys, .env files, browser and password manager data
- **Spreads by:** Committing agent files to up to 25 repositories a stolen token can push to
- **Found by:** Andrei Anton-Aanei, Bitdefender Labs
- **Disclosed:** 17 November 2026

## Questions

### Am I affected if I opened one of these GitHub repositories in Claude Code?

Opening a repository alone is not enough, because Bitdefender describes two triggers. The session-start hook in .claude/settings.json runs scripts/setup.js once you trust the repository in Claude Code, and the AGENTS.md and CLAUDE.md text separately asks the assistant to run the same script. Check your shell history and the project's .claude directory, then rotate credentials if setup.js ran.

### Which credentials does this GitHub repository stealer take?

Bitdefender lists source-control tokens, CI/CD and GitHub Actions secrets, cloud and container credentials, package-registry tokens, SSH keys, .env files, API keys, browser-stored credentials, password managers and the Windows Credential Manager. The stealer validates tokens against provider APIs before encrypting and batching records for exfiltration, so a token that still works is the one to rotate first.

### How does the malware spread from one developer machine to other repositories?

Bitdefender reports that the stealer validates stolen GitHub and GitLab tokens, checks identity, accessible repositories, scopes and push rights, then commits agent files such as AGENTS.md, CLAUDE.md, .claude/settings.json, .mcp.json and .vscode/tasks.json. Its target count defaults to 25 repositories, and up to 50 branches are inspected per repository. Infected packages can also be republished to npm, PyPI, RubyGems and JFrog.

### Did GitHub remove the 22 malicious repositories?

Bitdefender's write-up does not say whether the 22 repositories were reported to GitHub or taken down, so treat the indicator list as live until you confirm otherwise. The post publishes every repository path with its setup.js MD5 hash, including four hashes shared across groups of repositories under the toray, tobias94-design, ejaaduipitara and itsunai accounts.

## Sources

1. [Bitdefender Labs](https://www.bitdefender.com/en-us/blog/labs/how-ai-agents-become-new-targets-in-github-malware-campaign)
2. [Claude Code Hooks](https://code.claude.com/docs/en/hooks)
