Hackers Inject Malicious Code into Mistral AI Software Package

According to Decrypt, Microsoft’s Threat Intelligence team disclosed that attackers had injected malicious code into Mistral AI packages distributed via the PyPI platform. This malicious code automatically executes when developers use the packages on Linux systems, downloading and running a malicious file named <code>transformers.pyz</code> in the background—the filename deliberately mimics the widely used Hugging Face Transformers library to evade detection. Microsoft noted that the malware primarily steals developers’ login credentials and access tokens. It avoids execution on Russian-language systems and includes logic that can randomly delete files on devices located in Israel or Iran. This attack is linked to the “Shai-Hulud” supply-chain campaign launched in September. In response, Mistral stated that its investigation found the attack originated from compromised developer devices, and its corporate infrastructure was not breached.