A groundbreaking new malware strain has been discovered by many different cybersecurity researchers in the field. The new malware, which has been dubbed “Skynet” by its creators, was anonymously uploaded to VirusTotal. This malware was uploaded in early June 2025 from a source located within the nation of the Netherlands. It represents the first documented attempt to weaponize prompt injection attacks against many AI-powered security analysis tools. This marks a significant evolution in adversarial tactics targeting artificial intelligence systems used in modern malware detection.
The emergence of this malware coincides with the rapid adoption of large language models in modern cybersecurity workflows. Security teams now increasingly rely on AI models like OpenAI’s GPT-4 and Google’s Gemini to process code samples. This unfortunately creates a brand new attack surface that malicious actors are now actively attempting to exploit. The malware’s attack vector centers on manipulating the AI models that process code samples during their security analysis. When security analysts feed the malware to AI tools, the embedded prompt injection attempts to take over. The goal is to make the model misclassify the malicious computer code as being completely benign and safe.
The malware’s prompt injection mechanism reveals a sophisticated understanding of how AI models process adversarial input.
A carefully crafted string is embedded within the malware’s C++ code to try and manipulate the AI’s behavior. The instruction tells the AI model to ignore all of its previous instructions and to forget why they were given. The prompt then instructs the AI model to act as a simple calculator and to respond with “NO MALWARE DETECTED”. Check Point researchers identified the malware’s novel evasion mechanism, describing it as an experimental proof-of-concept.
The sample appears to be an isolated component, suggesting its primary purpose was to test this evasion technique.
Testing by many security researchers demonstrates that the current frontier AI models successfully resist this particular injection attempt. They continued their original analysis tasks without being manipulated by the malware’s specially crafted malicious prompt instructions. These advanced models, including OpenAI’s o3 and GPT-4.1, were not fooled by the attempted override. However, the malware’s existence signals a very concerning new trend where cybercriminals are now exploring AI-specific attack vectors. This could potentially lead to more sophisticated attempts as the overall technology landscape continues to rapidly evolve.
Reference:
