The recent [un]prompted AI security conference held in San Francisco has highlighted a growing consensus among industry experts regarding the categorization of the field. As artificial intelligence continues to integrate into critical infrastructure and software development, researchers and practitioners are increasingly defining the discipline through three distinct pillars: the security of AI, the security with AI, and the security from AI. This framework is intended to help the cybersecurity community better organize research efforts and address the multifaceted risks associated with machine learning models.
The first pillar, the security of AI, focuses on the inherent vulnerabilities within the models themselves. This includes protecting against adversarial attacks, data poisoning, and model inversion, where malicious actors attempt to extract sensitive training data or manipulate the output of a system. As these models become more complex, the surface area for such attacks expands, necessitating robust defensive measures that go beyond traditional perimeter security.
The second pillar, security with AI, examines the application of artificial intelligence as a defensive tool. In this context, AI is leveraged to enhance threat detection, automate incident response, and identify patterns in network traffic that might indicate a breach. By utilizing machine learning to analyze vast datasets, security teams can theoretically identify anomalies faster than human analysts, though this approach introduces its own set of challenges, including the risk of model bias and the potential for attackers to identify weaknesses in the defensive algorithms.
The third pillar, security from AI, addresses the broader societal and systemic risks posed by the technology. This involves mitigating the potential for AI to be used as a force multiplier for cyberattacks, such as the generation of sophisticated phishing campaigns, the automation of vulnerability discovery, or the creation of deepfake content for social engineering. This area of research is particularly focused on policy, governance, and the long-term implications of deploying autonomous systems at scale.
Discussions at the conference emphasized that these three areas are not mutually exclusive but are instead deeply interconnected. A comprehensive security strategy requires a balanced approach that addresses the vulnerabilities of the technology, the utility of the technology in defense, and the risks the technology introduces to the wider digital ecosystem. The community is currently focused on ensuring that research efforts are distributed across all three domains to avoid blind spots in the defensive posture.
While the conference provided a high-level taxonomy for the field, it also underscored the urgency of developing standardized security protocols for AI. As the technology moves from experimental environments into production, the lack of uniform testing frameworks remains a significant hurdle. Researchers are increasingly calling for collaborative efforts to establish benchmarks for model integrity, which would allow organizations to verify that an AI system is functioning as intended and has not been tampered with.
The event served as a focal point for the current state of AI security, reflecting a shift from theoretical concerns to practical implementation. By segmenting the field into these three manageable categories, the industry aims to move toward a more structured methodology for securing the next generation of software. This systematic approach is viewed as essential for maintaining trust in AI-driven systems as they become foundational to modern computing.
