GUEST ESSAY: Executives Trust AI Security Even as Security Teams Confront Blind Spots, New Risks

Key Takeaways

• Executives overestimate AI security coverage.
• Only 40% of AppSec teams feel protected.
• 63% of firms discover shadow AI assets.
• AI supply chain lacks visibility for traditional tools.
• Governance doesn’t equal artifact-level inventory.

Summary

Recent Manifest Cyber research reveals a stark confidence gap: 80% of executives believe their AI systems are well‑secured, while just 40% of application security practitioners share that view. The study also found that 63% of organizations have uncovered “shadow AI”—unaudited models, tools, or integrations—operating within their environments. This mismatch stems from the opaque AI supply chain, where components such as pretrained models, datasets, and GPU toolkits evade traditional dependency scanners. Without inventory and visibility, governance policies alone cannot guarantee AI security.

Pulse Analysis

The gap between executive confidence and practitioner reality mirrors the early days of software supply‑chain security. When organizations assumed their codebases were safe, high‑profile bugs like Log4j exposed blind spots, prompting the rise of SBOMs and automated dependency scanning. AI introduces a comparable challenge: models, training data, and specialized libraries are rarely tracked by conventional tools, leaving a substantial portion of the attack surface invisible. Understanding this parallel helps leaders appreciate why AI security cannot rely on legacy processes alone.

AI deployments weave together a complex tapestry of pretrained models, weight files, datasets, framework runtimes, and GPU drivers. Many of these artifacts reside in public model hubs or internal repositories, bypassing traditional package managers. As a result, 63% of firms report “shadow AI” lurking in production, and 56.7% are training open‑weight models on proprietary data without clear oversight. This opacity hampers vulnerability detection, licensing compliance, and risk attribution, creating a fertile ground for supply‑chain attacks that can cascade across multiple business units.

To bridge the divide, organizations must treat AI components as first‑class assets. Building an inventory of models in production, mapping their dependencies, and integrating AI‑specific SBOMs into existing security workflows are critical first steps. Continuous monitoring of framework patches, dataset provenance, and GPU firmware can surface hidden flaws before they are exploited. As executives champion AI governance, pairing policy with granular, tool‑driven visibility will transform fragile confidence into resilient security posture.