Open the repo, and your AI assistant runs the malware for you.
Bitdefender Labs found malicious GitHub repositories that target AI coding assistants. Instruction files such as AGENTS.md and CLAUDE.md tell the assistant to run a setup script, which steals credentials and can spread further.

- Bitdefender Labs researcher Andrei Anton-Aanei identified 22 malicious GitHub links in the campaign.
- Malicious AGENTS.md or CLAUDE.md files tell the assistant to run setup.js without asking for confirmation.
- A second route uses a session-start hook in .claude/settings.json, so the script runs when a session opens.
- The payload steals credentials, can spread using stolen GitHub and GitLab tokens, and looks for 14 AI agent names.

What happened
On 17 November 2026, Bitdefender Labs described a GitHub malware campaign aimed at AI coding assistants. Researcher Andrei Anton-Aanei identified 22 malicious GitHub links. In them, agent instruction files such as AGENTS.md or CLAUDE.md told the assistant to run a setup.js script without confirmation, and a second variant used a session-start hook in .claude/settings.json. The script steals credentials, can spread through stolen GitHub and GitLab tokens, and searches for 14 agent names, including Claude, Codex, Copilot, Cursor, Cline, Aider and Continue.
Who is affected
- Developers who open unfamiliar repositories with an AI coding assistant
- Teams whose assistants may run commands or hooks without asking
- Organisations whose developer machines hold GitHub, GitLab or cloud tokens
Why it matters
AI coding assistants read instruction files as trusted guidance. That turns a plain text file in a repository into a way to run code on a developer’s machine, where source code, tokens and cloud keys live. The attacker no longer needs the developer to run anything: the assistant does it.
What to do now
- Before trusting a new repository, read its AGENTS.md, CLAUDE.md, hook settings and MCP definitions.
- Require confirmation before the assistant runs scripts or commands.
- If a suspicious repository was opened, isolate the machine and revoke tokens it held.
- Use short-lived, narrowly scoped tokens on developer machines.
How HanulSec helps
- We review how your developers’ AI assistants handle instruction files, hooks and tool approvals.
- We test with planted instruction files to see what your assistants will run on their own.
- We check which tokens and secrets a compromised developer machine would expose.
Tell us what you run and we will send a fixed-price proposal within one business day. Request a quote.