The Data Dilemma: Navigating the AI Systems Matrix

Understanding AI security requires mapping the specific threat vectors that traditional cybersecurity frameworks were not built to address:

• Supply Chain Vulnerability poses a high risk where AI models can be compromised during training and development, embedding invisible biases that quietly influence downstream programmes, including strategic decisions. This threat vector indicates open-source dataset contamination, compromised defence-contractor pipelines, or unverified ingestion of sensor logs. Such as, when an AI detection system is used for automated target recognition (ATR), it can be compromised by open-source synthetic radar or satellite imagery. A small percentage of variation in target images or information, when introduced through specific levels of misinformation, may lead to miscalculation. It will lead to major disruptions in operations and overall assessment; for instance, civilian assets may appear to be high-value military targets.
• Decision-making is a challenge in which it is difficult to trace AI’s influence on strategic assessment, creating uncertainty in critical processes once a system is suspected of compromise. Specifically, military decision frameworks are the rules, data flows, and command structures commanders use to evaluate threats and act. Adding AI turns these systems into hybrid human-machine mechanisms. It includes three contending factors: velocity mismatch, in which AI processes battlefield data in milliseconds, forcing human commanders to either blindly trust the machine or lose tactical advantage; opacity, the deep learning models do not show why they reached a specific conclusion, leaving decision makers unable to verify if a target or risk assessment is correct; Volatility of AI models fail unpredictably when faced with data they were not trained to handle, leading to sudden, catastrophic errors. This can cause chaos in emergencies and crises.
• Interface Manipulation arises from unpredictable interactions between AI systems, letting attackers trigger malicious behaviour while the system continues to appear normal. Adversarial evasion alters operational data at the point of inference to deceive a deployed AI model without corrupting its underlying baseline data. This threat vector is associated with Electronic Warfare (EW), in which signal noise is injected into synthetic aperture radar (SAR) systems, or with deceptive acoustic profiles in submarine sonar processing. For example, an autonomous Unmanned Aerial Vehicle (UAV) uses computer vision for terrain navigation and threat identification. The tempering of physical modifications, such as specific geometric tape patterns on defence equipment or active infrared lighting pulses to a surface-to-air missile site. To a human operator, the target remains obvious, but the computer vision model fails to classify it due to high-dimensional (HD) feature distortion, blinding the autonomous platform’s target recognition system.
• Speed is the defining vector of the threat landscape. The world is engaged in an information and AI systems race in which AI vulnerabilities grow exponentially. AI-based systems can manipulate and plan malicious actions faster than human defenders can detect. This asymmetry makes such systems difficult to defend while giving attackers powerful new tools. Deepfake material, social engineering content, and algorithmic manipulation can now be generated at scale in minutes; catching them still requires careful, often manual analysis.
• The Governance-Accountability Gap is the counterproductive element of threat vectors. The most serious vulnerability is not technical; it is governance. Traditional security governance assumes clear, singular accountability resting with a human decision-maker. AI security instead requires frameworks built for distributed decision-making, where responsibility is shared among systems, developers, and end users. No amount of technical patching bridges that gap on its own. There is a pressing need to establish new frameworks for evaluating the operational value of AI system integration, and to ensure that human oversight remains substantive rather than merely nominal.