Securing the AI flank to provide trusted decision-making capabilities

Summary

Trustworthy AI in military settings demands more than ethical guidelines. It requires systems that remain predictable, verifiable, explainable, and resilient when adversaries actively target them. Mandy Satterwhite, Managing Director and Cyber Lead at Accenture Federal Services, stresses that command-and-control tools must deliver recommendations grounded in traceable, untampered data. Without clear data provenance or an intelligible derivation path, operators cannot act on AI outputs when lives depend on the outcome.

Commercial foundation models introduce distinct supply-chain exposures. Fine-tuning a frontier model imports whatever poisoned training data or compromised weights it carries, and adversaries have already attempted direct manipulation of model files. Satterwhite notes that oversized general-purpose models are often unnecessary for specialized defense tasks. Instead, organizations should right-size models to mission requirements and audit every component—from source data through weights—before integration into secure environments.

Traditional periodic penetration testing no longer matches the pace of AI-specific threats such as prompt injection or weight tampering. Continuous, automated validation embedded in CI/CD pipelines is required, with AI-driven red-teaming running alongside development. For national-security systems, real-time checks against controls expressed in NIST’s OSCAL language help ensure designs remain secure by construction rather than through late-stage remediation.

Security integrated from the outset accelerates rather than impedes deployment. When validation occurs throughout the build process, late redesigns are avoided and accreditation proceeds more rapidly. At the same time, human authority is preserved through explicit architectural guardrails, OODA-loop checkpoints, and kill switches. AI can surface verified courses of action from sensor data, yet a human operator evaluates those options at defined decision points before any action is authorized.

Why it matters

Directly addresses AI security, adversarial resilience, and supply-chain integrity for defense applications, highly relevant to Dutch and NATO defense professionals focused on ethical, secure AI deployment.