DIGITAL LIFE

Check Point research announces first documented case of AI-evading malware

The discovery, announced by Check Point Research, reveals a new category of cyberthreat designed to bypass AI-powered security systems. This is the first documented case of AI-evading malware that uses natural language text to manipulate defense tools.

The malware was designed to be classified as safe by large-scale language models (LLMs). This incident, despite its failure, marks the beginning of a new front in cybersecurity focused on exploiting AI tools themselves.

Analysis of the file, uploaded anonymously to the VirusTotal platform, revealed a string of text that sought to trick an AI model. The instruction, written in C++, asked the system to ignore previous commands and respond with the message “NO MALWARE DETECTED”.

While it also contains traditional tactics, such as TOR network components and sandbox evasion, the focus on AI manipulation is what sets this threat apart. Check Point’s detection system was able to identify both the malicious nature of the file and the prompt injection manipulation attempt, successfully neutralizing the attack.

AI-Evading Malware: A New Phase in the Evolution of Cyberattacks...The emergence of AI-evading malware marks a new race between attackers and defenders, similar to the one that occurred with the emergence of techniques to avoid detection in sandbox environments. As companies integrate artificial intelligence into their defenses, attackers are beginning to develop methods to exploit its weaknesses.

This development confirms that cybercriminals are not only adapting to new protection technologies, but are also actively studying them to turn them into attack vectors. The strategy is no longer just hiding the malware, but also actively deceiving the intelligence that is looking for it.

In short, the identification of this AI-evasion malware serves as a wake-up call for the industry. Even if its first attempt failed, its intent to manipulate AI systems has been proven. This event foreshadows a future where digital threats will become increasingly sophisticated, requiring AI-based security tools to constantly evolve to protect themselves.

https://research.checkpoint.com/2025/ai-evasion-prompt-injection/

mundophone