⚪ A new survey from HiddenLayer finds that 93% of organizations are using open-weight AI models sourced from public repositories, yet fewer than half consistently scan those models for tampering or malicious code before deployment. The 2026 AI Threat Landscape Report, drawn from responses from 500 security professionals, reveals a sector-wide security debt accumulating underneath rapid AI adoption.
The numbers illustrate the scale of the exposure. Thirty-one percent of respondents said they cannot confirm whether their organization has experienced an AI-related security incident. Seventy-six percent identified shadow AI as an active problem inside their environments. Models are being adopted faster than the security controls needed to govern them.
The five threats HiddenLayer identified
The report categorized the primary attack surfaces into five areas: data poisoning and backdoors embedded in training data, AI supply chain attacks targeting model distribution pipelines, prompt injection and guardrail bypass, memory and retrieval-augmented generation (RAG) poisoning, and model evasion techniques. Each exploits a different layer of the AI stack, and most organizations lack dedicated controls for more than one or two of them.
The structural problem
AI is now business-critical for most organizations, but the security investment required to protect AI systems has not kept pace. Scanning open-weight models before deployment is a basic hygiene practice. Doing it for fewer than half of deployments means the majority of organizations are running models they cannot fully verify. The same dynamics that created the open-source software supply chain crisis in 2020 and 2021 are replicating inside AI pipelines today.
What to prioritize
Start with model provenance verification: know where every model in production came from and whether it has been cryptographically verified by the source. Build a scanning pipeline for open-weight models before they reach any production environment. Treat AI model files with the same rigor applied to third-party code dependencies. HiddenLayer's report is a useful baseline for benchmarking your current controls against what the broader security community is actually doing.
Gigia Tsiklauri is a Security Architect and founder of Infosec.ge. Get in touch if you want help developing an AI model security assessment for your organization.