A Narrative Review of Artificial Intelligence Vulnerabilities in Cybersecurity

Artificial intelligence is used in malware classification traffic analysis software development and security decision support. Yet conventional accuracy metrics do not capture the attack surfaces introduced by these applications. This narrative review critically examines 35 purposively selected studies published between 2020 and 2026. Primary records were verified against publisher proceedings institutional and author sources. The corpus comprises 33 peer-reviewed publications and two identified preprints covering evasion poisoning privacy leakage model extraction language-model injection and evaluation assurance. The analysis distinguishes findings from security applications from mechanisms established in general machine -learning and agent benchmarks. Across these settings vulnerability depends on an attacker’s access whether manipulations are valid within the application domain, and whether an erroneous output can cross an authorization boundary. Malware attacks that preserve functionality and traffic attacks that respect domain constraints offer stronger evidence of feasibility than perturbations confined to feature space. Retrieval and persistent memory extend the scope for poisoning beyond training. Agent benchmarks, in turn, highlight the need to assess authorized task completion separately from unauthorized actions. Contrasting findings from developer studies also indicate that insecure model output is not a uniform measure of harm to users. The proposed lifecycle assurance framework brings together data provenance realistic adaptive evaluation restricted privileges and monitoring over time. The purposive sample and heterogeneous outcomes preclude pooled estimates of attack success or claim about prevalence. Deployment decisions therefore require evidence spanning the full decision and action pipeline and robustness claims must remain limited to the conditions evaluated.