Recently, a typical case of using large language models (LLMs) to assist in developing malicious software has been exposed in the field of cybersecurity. According to security researcher Sibi Moosa's monitoring, an attacker with the online name "mousie-5212-super-formatter" is accused of using Anthropic's AI model Claude to write malicious code and extensively pollute the npm package manager ecosystem. In a short period, the attacker pushed over 670 malicious software packages into the npm repository, and the level of automation and speed of generation have raised high concerns across the industry.

claude, Claude

The core of this attack lies in using AI to significantly lower the barrier for writing malicious code. The polluted npm packages are designed to steal developers' sensitive credentials (such as npm tokens, GitHub tokens) and source code from internal private GitHub repositories. The attacker uses Claude to generate logically sound stealing scripts and uploads the obtained data to their controlled repositories. This incident reveals that while generative AI enhances development efficiency, it is also becoming a "multiplier" for attackers to improve attack efficiency and automation levels.

Experts point out that using AI models for automated package pollution and code theft marks a new intelligent phase in supply chain attacks. Traditional signature-based defense mechanisms struggle to deal with these highly variable and deceptive malicious carriers generated by AI. As AI programming assistance tools become more widespread, how to prevent models from being misused for vulnerability exploitation and malicious development has become an urgent issue in AI security governance.