AI hallucinations put critical infrastructure at risk study shows

AI hallucin­ations are introducing serious security risks by delivering highly confident yet incorrect outputs that mislead critical infrastructure decision‑making. The piece cites a 2025 Artificial Analysis AA-Omniscience benchmark which found that, across 40 AI models, all but four were more likely to give a confident wrong answer than a correct one on difficult questions.

It emphasises that as AI plays a larger role in cybersecurity, every AI‑generated response should be treated as a potential vulnerability until a human verifies it. The article outlines three main impacts: missed threats, fabricated threats and incorrect remediation, with examples such as misclassifying normal traffic as malicious or, conversely, failing to flag zero‑day techniques.

It also stresses governance measures like insisting on human review before actions, auditing training data and enforcing least‑privilege access, and notes Keeper Security’s role in centralising identity controls to prevent harm when AI guidance proves incorrect. According to Keeper Security, the contributed piece was written by Ashley D’Andrea.