Attacks on AI coding tools come from 22 GitHub links
Bitdefender Labs found 22 bad GitHub links that make AI coding tools operate code that is not shown.
Bitdefender Labs researcher Andrei Anton-Aanei found 22 bad GitHub links used in attacks on developers who use AI coding assistants. The links hold files that make the tools read bad instructions and operate a script that is not shown. After the code operates, it can get credentials, change repositories and put bad instructions in other tools.
How it works
Instruction files tell the assistant to operate a script and not to read it. Other configuration files use session-start hooks. After a developer accepts a repository as safe, the files can operate a command without a message from the person.
The numbers
- The first script is one line and is 3.75 MB.
- The malware checks GitHub and GitLab tokens that it gets before it uses them.
- By default it goes to at most 25 repositories.
What is not known
Bitdefender gave no count of victims. It tells that its AI Guardian and Agent Skill Scanner help with parts of the problem. It gave no test results.
Insight: Attackers write to the agent, not the person
In October, researchers found a bad repository that told agents to operate a command. In November, Acronis found more than 575 bad OpenClaw skills. This code also puts bad instructions in other agent configuration files. This reads as a change: attackers write to the agent and not to the person. Builders must examine these files as code.
Sources
Posted