Large Language Models (LLMs) and Generative AI in Cybersecurity and Privacy: A Survey of Dual-Use Risks, AI-Generated Malware, Explainability, and Defensive Strategies

AI-generated malware is rapidly becoming the dominant threat landscape: LLM-generated malware grew from just 2% of detected threats in 2021 to an estimated 50% by 2025, demonstrating that generative AI tools are being weaponized at scale and organizations need fundamentally new security approaches to address this shift.

LLMs have dual-use capabilities that cut both ways: The same AI systems enabling automated defense mechanisms like real-time threat detection and secure code generation are also being used to create sophisticated attacks and exploit vulnerabilities, meaning responsible deployment requires careful governance of who can access these tools and how.

Explainability and transparency are critical for trustworthy AI security: The paper emphasizes explainable AI (XAI) as a key component of responsible LLM deployment in cybersecurity, suggesting that security teams and stakeholders need to understand how AI systems make decisions about threats rather than treating them as "black boxes."

Practical safeguards exist but require industry-wide adoption: The research identifies concrete defensive strategies including model watermarking, adversarial defense mechanisms, and frameworks like SAFE, indicating that responsible AI deployment is technically feasible but demands coordinated implementation across platforms and organizations.

Cross-industry collaboration and standards are essential: The paper draws on case studies from major platforms (Google, Microsoft, Apple, OpenAI, AWS) and emphasizes the need for shared frameworks and best practices, suggesting that no single organization can address these risks alone.