Hackers used an AI code tool settings file to start a backdoor
Google tells that a suspected North Korea group changed a settings file, and Claude Code then downloaded a backdoor.
Google tells in a new report that a suspected North Korea group attacked cryptocurrency firms in North America and Europe. In 2026 the group changed a settings file in the internal code stores of the firms that it attacked. When a user started Claude Code, the file made it download a backdoor, which is hidden remote access.
How it worked
The group made fake recruiter profiles with AI pictures. It sent targets to fake meeting sites that told them to start a script. Then it moved into the code stores. It also gave access tokens, which are digital keys, the name Microsoft Copilot to hide its moves.
Where AI helps less
Google tells that AI tools make much noise in logs, and models can make mistakes about facts or targets. Google expects spy groups that must have stealth to use them less. Google is sure with high confidence that crime groups will keep them, because more targets make up for less success.
What is not known
Google gives no figure of loss. The connection to North Korea is a suspicion.
Insight: The settings file of an AI tool is a new door
Google found a settings file that started a backdoor. Researchers showed an attack that adds one tool step to an AI agent. It sends data out in 62.64% to 100% of tests. This reads as a plan to use the trusted setup around the agent and not the model. Tool settings are part of your security border. Control them as you control code.
Sources
Posted