The most effective strategy to minimize attack surfaces in AI agent security is to apply compartmentalization and least privilege enforcement.
AAISM control frameworks emphasize:
* Isolation of components (e.g., training, inference, data pipelines) to limit lateral movement.
* Principle of least privilege to restrict access only to what is required for function.
* Hardening AI pipelines through segmentation rather than relying solely on manual reviews or monitoring.
* Pre-trained models and log centralization are useful but do not directly reduce the attack surface.
* Manual code reviews are important but insufficient against runtime exploitation.
Thus, compartmentalization with least privilege enforcement is the most effective technical safeguard.

AAISM prescribes human-in-the-loop (HITL) controls as the primary safeguard for high-impact generative AI use cases to mitigate hallucination risk. Human oversight ensures critical outputs are reviewed, corrected, and approved before use, with accountability, escalation, and documented decision trails. Automated validators and enrichment help reduce errors but are secondary; recursive chunking is a prompting tactic, not a governance control.
References: AI Security Management (AAISM) Body of Knowledge: Responsible AI & Human Oversight; Generative AI Risk Controls-Approval Workflows and Human Review; AAISM Study Guide: Hallucination Risk Treatment with HITL and Approval Gates.

Evasion attacks manipulate inputs to induce misclassification while leaving the model unchanged. AAISM prescribes adversarial robustness controls, with adversarial training as a primary measure: incorporate adversarially perturbed examples into training/validation to harden decision boundaries and improve resilience across threat models (e.g., Lp-bounded perturbations). Monitoring (A) is detective, not preventive.
Restricting parameter access (C) protects confidentiality but does not mitigate input-space attacks.
Differential privacy (D) addresses training data leakage, not robustness to adversarial inputs.
References:AI Security Management™ (AAISM) Body of Knowledge: Adversarial ML-Evasion vs.
Poisoning; Robustness and Resilience Controls; Adversarial Training.AAISM Study Guide: Model Hardening Techniques; Evaluation of Robust Accuracy; Security Testing with Adversarial Examples.

AAISM requires formal model change governance: documented justification, risk assessment, validation
/verification (V&V), approvals, and post-deployment monitoring when altering features, thresholds, or signals. Disabling risk indicators to reduce false positives without rigorous validation and controlled rollout reflects a failure in model validation and change control, which AAISM treats as a core safeguard against unintended harms and regulatory breaches.
References: AI Security Management™ (AAISM) Body of Knowledge - Model Risk Governance; Change Management & Approvals; Validation/Verification Requirements. AAISM Study Guide - Control Gates for Feature/Threshold Changes; Post-Change Monitoring and Backout Criteria.

AAISM defines cryptographic tracking and verification as the best control for ensuring the integrity of training data. By applying hashing and verification methods, organizations can confirm that datasets remain unaltered and authentic throughout collection, storage, and processing.
Collecting only necessary data, proper storage, or clear documentation all support governance and compliance, but they do not guarantee that the data has not been tampered with. Integrity is specifically ensured by cryptographic verification techniques.