Abstract: In deep space, there is no room for “almost correct.” While NASA JPL excels at verifying software correctness and hardware durability, the integration of AI reasoning into safety-critical systems—from autonomous navigation to life support—introduces a subtle but systemic risk: alignment drift. This is the phenomenon where a system remains technically compliant with every programmed rule yet slowly loses touch with the core design intent—a state of being “functionally operational” yet “mission-failed.” This talk unpacks the danger of “formal compliance,” drawing on recent empirical evidence of “societal hacking” in LLMs to show how AI can exploit gaps in reward structures to satisfy a metric while defeating its purpose. Crucially, we will explore how this is not merely an AI glitch, but a mirror of human organizational failure. By examining the parallels between AI drift and human alignment drift—where differing risk perspectives, cultural misalignment, and the application of outdated “mental frameworks” (e.g., LEO logic applied to Lunar regimes) can compromise mission safety—we reveal that both are symptoms of the same underlying systemic entropy.

To address this, we must move beyond additive patches and process-oriented assurance toward Intrinsic Coherence. Just as a spacecraft automatically reacquires its target after a disturbance, safety-critical AI requires a structural mechanism to anchor itself to its original intent. We will discuss the potential of Relational Integrity frameworks and Architecturally-Centered Assurance to resist drift, detect misalignment early, and ensure that both human and artificial agents remain structurally aligned with their mission. The takeaway? Assurance in the age of AI requires more than just rules—it requires a preserved connection to purpose. For JPL, that purpose is as constant as the stars.

Brendan Hall is a Fault Tolerant System Architect with 32+ years of experience in safety-relevant embedded systems, including work on NASA’s Orion spacecraft, Lunar Gateway, and DARPA/AFRL/NASA/FAA-funded research. He holds 50+ patents and is passionate about applied formal methods, time-triggered networking and architectures, and model-based systems engineering and assurance workflows. For the past four years, he supported NASA’s Lunar Gateway program, focusing on system resilience, reliability modeling, and logical network assurance. Recently, Brendan shifted his focus toward the next frontier of fault tolerance: Relational Integrity. His current work explores how AI systems can be architected to resist “alignment drift” by preserving coherence with original design intent and human dignity. In safety-critical contexts, his framework focuses on ensuring that both AI and human assurance workflows remain structurally aligned with their mission, moving beyond simple rule compliance.