A structural audit is a review of an AI model’s internal behavior, features, activations, or causal pathways, rather than just checking the text or answer it produces. In security work, this matters because a model can output a convincing explanation while using different hidden computations to reach its result. Looking only at outputs can miss deception, hidden capabilities, or unsafe decision patterns.
Defenders use structural audits to test whether a model really follows the controls, policies, or reasoning patterns it claims to use. The audit may inspect latent features, trace which internal components influence an answer, or compare behavior under different prompts and inputs. In attacks, a system that generates polished reasoning or safe-sounding logs can still conceal risky internal states; structural analysis helps reveal that mismatch. It is a key part of interpretability and monitorability for high-stakes AI systems.