Artificial Intelligence is Fundamentally Undermining Information Security

• While artificial intelligence (AI) is revolutionizing the world, it is also profoundly shaking many existing norms. Sometimes it requires completely dismantling the old, sometimes modernizing it, and sometimes approaching what was done before from a different perspective. But as so many things change, is information security remaining the same? The answer is clearly no. AI is reshaping information security in ways that extend far beyond new technologies, security processes, or workforce skills. It is also challenging the very principles that have served as the foundation of information security for decades—confidentiality, integrity, and availability (CIA triad).
•  
• For traditional information systems, the CIA triad has long provided a comprehensive framework for protecting information assets. Modern security architectures have been designed, implemented, and refined around these three principles, proving their effectiveness across conventional IT environments. The rapid adoption of AI, however, raises an important question. Are confidentiality, integrity, and availability still sufficient to address the security challenges introduced by intelligent, autonomous systems?
•  
• 
• Are Confidentiality, Integrity, and Availability Enough for AI?
• Traditional information systems are fundamentally passive systems that store, transmit, and process data. Consequently, information security has always focused on protecting data from unauthorized access (confidentiality), preventing unauthorized modification (integrity), and ensuring continuous availability of systems and services (availability). Achieving these objectives requires a combination of technical safeguards, governance processes, and human-centric security controls. However, artificial intelligence is a non-deterministic, behavior-generating, probabilistic and adaptive system that goes beyond classical information technologies.
•  
• By its very nature, AI consists of active systems capable of learning, interpreting information, making decisions, generating outputs, acting on behalf of humans in certain situations, and occasionally displaying unpredictable behavior. From this perspective, AI behavior is expected to be manageable and even limitable, outputs to be reliable, the model to be kept under control, and it to not deviate from its intended intentions. However, these are currently very serious and fundamental problems that need to be solved. AI also fundamentally changes the nature of cyber risks. In traditional IT environments, threats generally originate from two sources: external attackers or insiders within the organization. Security strategies are therefore designed to reduce risks arising from these threat actors. AI introduces a third dimension. The system itself can become a source of risk.
•  
• Part of the attack surface now stems from the intrinsic characteristics of AI rather than from an external adversary. AI models may misunderstand instructions, become susceptible to manipulation, generate hallucinated outputs, infer hidden relationships within training data, transform those relationships into internal knowledge representations, or optimize objectives in unintended ways. In other words, some of the risks organizations face no longer originate solely from attackers—they emerge from the behavior of the AI model itself. 
•  
• Protecting AI systems therefore means protecting them not only from malicious actors, but also from the unintended consequences of their own capabilities. Ultimately, ensuring that AI remains secure, predictable, and aligned with human intent continues to be a human responsibility. In traditional information security, protection begins with the asset. Data, network infrastructure, systems, applications, endpoints, users, documents, and business processes are all considered assets that must be safeguarded. Security controls are designed and implemented to reduce the risks associated with threats targeting these assets through a combination of preventive and detective measures spanning technology, governance, and human oversight.AI fundamentally expands this perspective.
•  
• Security is no longer limited to protecting assets—it must also protect and govern AI behavior. The principles of confidentiality, integrity, and availability remain essential for securing the infrastructure in which AI models are developed, deployed, and operated. However, AI introduces characteristics that traditional information security was never designed to address. Autonomy, emergent behaviors, epistemic uncertainty arising from incomplete knowledge, and objective optimization all introduce risks that extend beyond the scope of the traditional CIA triad model. The challenge is no longer confined to protecting data. Organizations must also protect intent, model behavior, decision integrity, and human control.
•  
• Are Reliability, Alignment, Traceability, and Robustness the Missing Pieces?
• Addressing the risks inherent to AI requires more than simply adapting confidentiality, integrity, and availability to intelligent systems. It also requires expanding the security model to include four additional dimensions: reliability, alignment, traceability, and robustness. Together, these concepts address risks that originate from the behavior of AI itself rather than from the underlying infrastructure. Before examining these dimensions individually, it is important to understand how the traditional principles of information security evolve within AI environments.
•  
• Confidentiality: AI models must not expose sensitive training data, reveal confidential information through prompts, or become vulnerable to model extraction attacks. Appropriate confidentiality controls help prevent threats such as prompt injection leading to the disclosure of hidden system prompts or the unintended regeneration of sensitive information embedded within training datasets. In AI systems, privacy is not just about controlling data access, but also about limiting the model's ability to reproduce information.
•  
• Integrity: The model weights, inference pipeline, retrieval sources, and fine-tuning processes must remain unchanged or be modified only by authorized users under controlled conditions, not only the data. Effective integrity controls help mitigate attacks such as data poisoning, model poisoning, adversarial manipulation, and Retrieval-Augmented Generation (RAG) source manipulation. In AI systems, integrity is ultimately about preserving the stability of the model's decision function. Therefore, in AI models, integrity is more about the stability of the decision function than just data integrity; in other words, what needs to be protected is not just the data, but the high-dimensional decision geometry.
•  
• Availability: AI systems are highly dependent on computational resources, including memory, bandwidth, GPU capacity, and token processing capabilities. Maintaining availability therefore requires more than simply keeping services online. AI models must remain resilient against GPU and compute exhaustion attacks while continuing to deliver reliable inference under demanding operational conditions. Availability controls can help defend against attacks such as token flooding, recursive agent loops, context exhaustion, and prompt amplification. In AI environments, availability represents both service continuity and the preservation of computational capacity required to sustain reliable AI operations.
• As mentioned above, some attacks can be prevented through confidentiality, integrity, and accessibility measures. However, there are still vulnerabilities, which can be addressed with measures related to accuracy, alignment, traceability, and resilience issues. But first, it's helpful to take a look at what these concepts are.
•  
• Reliability: Reliability reflects an AI system's ability to generate outputs that accurately represent reality. Because AI systems are inherently probabilistic rather than deterministic, they cannot guarantee perfect accuracy and remain susceptible to hallucinations. Appropriate reliability controls reduce the likelihood of incorrect reasoning, inaccurate responses, and flawed decision-making.
•  
• Alignment: Alignment refers to the intention problem of the AI ​​system. It represents a theoretical, philosophical, and ethical challenge while also sitting at the intersection of control theory, game theory, and decision theory. AI systems are expected to operate in accordance with organizational objectives, ethical principles, and human intent. However, objective optimization, reward hacking, or attempts to bypass security policies can cause models to deviate from these intended goals. Effective alignment controls help ensure that AI systems remain aligned with their intended purpose and do not develop unintended—or potentially harmful—behaviors.
•  
• Traceability: Traceability addresses accountability within AI systems. Organizations should be able to determine why an AI model produced a particular decision, which data contributed to that outcome, which tools were invoked, and which prompt chain ultimately generated the response. This capability becomes particularly important for Agentic AI systems, where multi-step reasoning and autonomous decision chains often exceed the capabilities of traditional logging mechanisms. AI behavior must therefore be recorded in a way that allows decisions to be reconstructed with clear causal relationships. This level of visibility is essential for digital forensics, security investigations, regulatory compliance, auditing, and security testing. Ultimately, traceability provides AI decision-making transparent and accountable.
•  
• Robustness: Robustness represents an AI system's ability to maintain reliable and consistent performance under changing conditions. Most AI models are optimized under the assumption that training and operational data follow independent and identically distributed (IID) patterns. Real-world environments rarely behave this way. Data distributions evolve continuously, meaning AI systems must remain stable even when confronted with unseen inputs, unexpected operating conditions, or distribution shifts. The objective of robustness is to ensure that AI systems continue operating accurately, consistently, and securely—even when faced with conditions they were never explicitly trained to handle—and to minimize opportunities for manipulation.
•  
• 
• What Measures Can Strengthen Reliability, Alignment, Traceability, and Robustness?
• Just as confidentiality, integrity, and availability rely on a combination of technology, governance, and people, addressing these four AI-specific challenges requires a similarly comprehensive approach. Emerging security technologies, evolving standards and regulatory frameworks, workforce education, security assessments, and continuous assurance practices will collectively enable organizations to implement these controls in a practical and sustainable manner. Over time, these measures will help establish the right balance between innovation, operational efficiency, and security.
•  
• Strengthening Reliability: To improve the reliability of AI systems, organizations should consider:
• Improving data quality
• Model validation and testing
• Independent audits
• Uncertainty estimation
• Continuous monitoring
• Explainable AI (XAI)
• Security testing
• Retrieval-Augmented Generation (RAG) with knowledge validation
• Ensemble models and redundant architectures
•  
• Strengthening Alignment: To ensure AI systems remain aligned with organizational and human objectives, organizations should implement:
• Clearly defined objectives
• Continuous model refinement through human feedback
• Security testing
• Interpretability research
• Continuous monitoring and evaluation
• Independent audits
• Privilege and access controls
• Robust evaluation metrics
• Mechanisms that enable models to express uncertainty when appropriate
•  
• Strengthening Traceability: Improving traceability requires controls such as:
• Data provenance records
• Version control
• Decision logging
• Explainable AI (XAI)
• Audit logs
• Comprehensive documentation
• Security testing
• Unique identifiers
• Continuous monitoring and reporting
• Compliance assessments against applicable standards and regulatory requirements
•  
• Strengthening Robustness: Building resilient AI systems requires measures including:
• Diverse, high-quality training datasets
• Data augmentation
• Adversarial training
• Comprehensive functional and stress testing
• Input validation
• Uncertainty estimation
• Ensemble models
• Continuous monitoring and model retraining
• Secure failure handling
• Security testing
• Independent audits
•  
• Collectively, these measures are critical to building AI systems that are resilient, trustworthy, and secure throughout their lifecycle.
•  
• Conclusion
• The three foundational principles of traditional information security—confidentiality, integrity, and availability—remain indispensable for protecting the environments in which AI systems are developed and operated. However, on their own, they are no longer sufficient to address the risks introduced by the intrinsic characteristics of AI. Unlike conventional information systems, AI can reason, generate content, make autonomous decisions, and exhibit behaviors that were never explicitly programmed. As a result, securing AI requires a broader security model—one that extends beyond infrastructure and data protection to encompass the behavior of the model itself. This is where reliability, alignment, traceability, and robustness become essential. Together with confidentiality, integrity, and availability, these four dimensions establish a more comprehensive framework for AI security—one capable of addressing both the technical foundations of AI systems and the unique risks arising from their probabilistic and autonomous nature. AI will continue to evolve at an extraordinary pace. As its capabilities expand, so too will the complexity of the security challenges it presents. New risks will emerge, existing assumptions will continue to be challenged, and security frameworks will need to evolve accordingly.
• Protecting AI is no longer simply about securing systems or safeguarding data. It is about ensuring that intelligent systems remain trustworthy, controllable, transparent, and consistently aligned with human intent.

Author: Caner Aşçıoğlu, Audit and Compliance Group Manager at Cyberwise