Marcelo Delima | Senior Manager, Global Solutions Marketing
More About This Author >
Marcelo Delima | Senior Manager, Global Solutions Marketing
More About This Author >
Manufacturers have always depended on automation. From robotics on the factory floor to predictive maintenance, connected supply chains and digital twins, the industry has long understood the value of machines that can work faster, more consistently and more efficiently than humans alone. But the next wave of automation is different. Agentic AI systems do not simply analyze data or generate recommendations. They can reason, invoke tools, make decisions and act autonomously across business and operational systems. In manufacturing, where operational uptime, intellectual property, supply chain continuity and worker safety are all mission critical, that shift has profound data security implications.
The 2026 Thales Data Threat Report: Manufacturing Edition shows that manufacturing organizations are rapidly confronting this reality. Two-thirds of manufacturing respondents (67%) cited rapid change in AI ecosystems as their top AI security risk. That is not surprising. AI platforms, models, APIs, Model Context Protocol (MCP) servers and agent frameworks are evolving faster than many governance models, security tools and operational processes can adapt. The risk is no longer only that a human user might make a mistake with a chatbot. It is that an autonomous or semi-autonomous agent could access sensitive data, trigger workflows, interact with other systems and amplify mistakes at machine speed.
The Manufacturing Edition draws on a survey of 417 security and IT executives at durable and non-durable manufacturers in 20 countries, conducted by S&P Global 451 Research as part of the 3,120-respondent 2026 Thales Data Threat Report. Here are the five data security risks it surfaces for the agentic era.
Key Takeaways: Manufacturing Cybersecurity in the Agentic AI Era
- 67% of manufacturers rank rapid change in AI ecosystems as their top AI security risk.
- Only 31% have complete knowledge of where their data is stored, and only 37% can classify all of it.
- Just 42% of manufacturers' sensitive cloud data is encrypted, down from 57% in 2025.
- 57% report rising credential theft and misappropriated secrets; cloud storage (36%), cloud-delivered SaaS applications (29%) and cloud management infrastructure (26%) are the top attack targets.
- 83% run five or more data protection tools, an average of eight, yet only 35% are confident they understand their data security tools.
- 59% name harvest now, decrypt later as their top quantum concern, and 53% are refactoring application and data architectures to meet sovereignty objectives.
Agentic AI Is the New Insider Threat in Manufacturing
In many ways, AI is becoming the new insider threat. Traditional insider risk typically focuses on employees, contractors or compromised identities with legitimate access to systems. Agentic AI changes that model. An AI agent may be granted access to data repositories, engineering systems, customer records, supplier portals, code repositories, MES and ERP platforms or operational dashboards so it can complete a task. If that agent is misconfigured, manipulated or insufficiently governed, it may expose data, take unauthorized action or make decisions no human intended. The agent itself may not be malicious, but its access, autonomy and speed can turn it into a powerful pathway for risk.
Sensitive data disclosure is already one of the most visible warning signs. According to the 2026 Thales Data Threat Report, Manufacturing Edition, 59% of manufacturing respondents said sensitive data disclosure attacks against AI or LLM environments are increasing. Manufacturing organizations hold enormous volumes of high-value information: product designs, production data, supplier terms, customer information, quality data, maintenance records, software code and operational telemetry. As agentic systems gain access to more of this information, the question becomes less about whether AI can improve productivity and more about whether organizations can control what AI can see, use, infer and share.
Data Visibility: Only 31% of Manufacturers Know Where Their Data Is Stored
That control starts with visibility. Yet only 31% of manufacturing organizations said they have complete knowledge of where their data is stored, and only 37% said they can classify all their data. You can’t protect what you can’t see. Lack of visibility leads to poor security outcomes: only 42% of manufacturers’ sensitive cloud data is encrypted, down from 57% in 2025. These gaps matter because AI agents are very good at finding and using information. If security teams cannot locate, classify and protect sensitive data, agents may discover it anyway. The result can be unintended disclosure, over-permissioned access or sensitive information being pulled into workflows where it does not belong. In an agentic environment, unknown data does not stay hidden. It becomes discoverable, reusable and potentially exposed.
Add yCloud and Credential Theft: Where Manufacturing Cyber Attacks Concentrateour headline here
Cloud environments are especially important because they are now central to both AI innovation and attack activity. Manufacturing respondents identified cloud storage (36%), cloud-delivered SaaS applications (29%) and cloud management infrastructure (26%) as the top attack targets. Credential theft is the leading attack technique against cloud infrastructure, and 57% of manufacturers report that credential theft and misappropriated secrets are increasing. This creates a dangerous combination: AI agents require access to data and tools, while attackers increasingly seek the credentials and secrets that make that access possible. If an attacker compromises credentials used by an AI workflow, the blast radius could extend far beyond a single application or user account.
Tool Sprawl: 83% of Manufacturers Run Five or More Data Protection Tools
Complexity makes the challenge harder. The report found that 83% of manufacturing organizations use five or more data protection tools, with an average of eight tools for data protection and monitoring, and 42% run five or more key management systems. At the same time, only 35% of respondents said they are confident in their understanding of data security tools. Tool sprawl creates gaps, inconsistent policies and operational confusion, and every extra console is another place an agent’s permissions can drift out of view.
Beyond AI: Digital Sovereignty and the Harvest Now, Decrypt Later Quantum Risk
Manufacturers also need to look beyond today’s AI risks. Digital sovereignty is becoming more important as organizations rethink where data lives, who can access it and how applications can be refactored to meet regional requirements. The report found that 53% of manufacturing respondents are reworking and refactoring application and data architectures as their main focus in achieving sovereignty objectives. Quantum risk is also moving from theory to planning. Future decryption of today’s encrypted data, often described as harvest now, decrypt later, was the top quantum concern for manufacturing respondents, cited by 59%, and 61% are already prototyping or evaluating post-quantum cryptography (PQC) algorithms. These issues may seem separate from agentic AI, but they share a common foundation: organizations must know their data, protect it cryptographically and govern its use across increasingly dynamic environments.
What Manufacturing Security Leaders Should Do Next
For manufacturing leaders, the path forward is not to slow AI adoption. The opportunity is too significant. Agentic AI can help improve uptime, accelerate engineering, streamline supply chain operations and support faster decision-making. But trusted AI requires trusted data security. Five priorities follow directly from the findings:
- Discover and classify all sensitive data before agents do. With only 31% knowing where data is stored and 37% able to classify it, continuous data security posture management (DSPM) is the precondition for safe agent access.
- Close the encryption gap. Encrypting 42% of sensitive cloud data leaves most of it readable to any over-permissioned agent or stolen credential; pair encryption with centralized key management so coverage is measurable.
- Treat agent credentials and secrets as privileged identities. With 57% seeing credential theft rise, apply least-privilege, short-lived credentials and secrets management to every AI workflow, not just human users.
- Consolidate the tool stack. Eight data protection tools and five or more key management systems cannot enforce one policy; reduce consoles so controls apply consistently across hybrid and OT-adjacent environments.
- Monitor AI workflows continuously and plan for sovereignty and PQC together. Log what agents access, infer and share, and fold post-quantum migration and data residency into the same data-governance program rather than running three separate projects.
Agentic AI Security in Manufacturing: FAQs
What is agentic AI in manufacturing, and why is it a security risk?
Agentic AI refers to AI systems that can reason, invoke tools, make decisions and act across business and operational systems without step-by-step human direction. In manufacturing, agents may touch engineering data, supplier portals, MES and ERP platforms and production telemetry. The risk is that an over-permissioned, misconfigured or manipulated agent can expose data or take unintended action at machine speed.
Why does the 2026 Thales Data Threat Report call AI the new insider threat for manufacturers?
Because AI agents operate with legitimate access, like an insider. 67% of manufacturers rank rapid change in AI ecosystems as their top AI security risk, and 59% say sensitive data disclosure attacks against AI or LLM environments are increasing. An agent does not need malicious intent to cause harm; broad access combined with poorly classified data is enough.
What are the biggest data security risks facing manufacturers in 2026?
The Manufacturing Edition points to five: rapid change in AI ecosystems (67%), limited data visibility (only 31% know where all their data is stored), weak cloud encryption (42% of sensitive cloud data encrypted), rising credential theft and misappropriated secrets (57%), and tool sprawl (83% run five or more data protection tools).
How much of manufacturers’ sensitive cloud data is encrypted?
Only 42% of manufacturers’ sensitive cloud data is encrypted, according to the 2026 Thales Data Threat Report, Manufacturing Edition, down from 57% in 2025. Cloud storage (36%), cloud-delivered SaaS applications (29%) and cloud management infrastructure (26%) are the top attack targets, so the unencrypted majority sits exactly where attackers concentrate.
What does “harvest now, decrypt later” mean for manufacturers?
Harvest now, decrypt later describes attackers stealing encrypted data today so they can decrypt it once quantum computers can break current algorithms. It is the top quantum concern for 59% of manufacturing respondents, and 61% are already prototyping or evaluating post-quantum cryptography (PQC) algorithms. Long-lived intellectual property such as product designs makes manufacturers especially exposed.
Download the 2026 Thales Manufacturing Data Threat Report for the full findings, including the breakdowns behind the figures above, or attend our webinar with Oracle and S&P Global, “2026 Manufacturing Data Threat Report,” for expert insights from 451 Research, Oracle and Thales.