Estimated reading time: 8 minutes · Last updated:
Netskope Threat Labs says manufacturing AI security risks are climbing as fast as the tools themselves. In its November 2026 sector report, the share of manufacturing users actively working with AI applications went from 40% to 67% in a year, and use of company-managed AI apps doubled from 35% to 70%. Exposure grew alongside adoption. Regulated data made up 57% of AI-related data policy violations, non-human entities connecting to remote Model Context Protocol (MCP) servers rose 490%, and malicious AI link interactions hit 200 per 100,000 users in March 2026. The figures come from anonymised telemetry on Netskope One customers gathered between 1 June 2025 and 31 August 2026. Netskope also sells the controls it recommends, which is worth keeping in mind.
Security teams should inspect links returned by AI applications the same way they inspect links in search results and email, before users or agents follow them.
Netskope Threat Labs, manufacturing report, November 2026
Key takeaways
- Managed AI doubled: Netskope Threat Labs found company-managed AI app use in manufacturing went from 35% to 70% in a year, while active AI users rose from 40% to 67%.
- Copilot leads: Microsoft Copilot appears at 90% of manufacturing organisations in Netskope's data, narrowly ahead of Claude Platform at 89%, with ChatGPT at 76%.
- Regulated data dominates: Regulated data accounted for 57% of AI-related data policy violations in the sector, against 19% for passwords and keys, 16% for intellectual property and 9% for source code.
- Agent traffic surges: Non-human entities reaching remote MCP servers increased 490%, and MCP events grew 600%, over Netskope's measurement period.
- AI links weaponised: Interactions with malicious links returned by AI tools peaked at 200 per 100,000 manufacturing users in March 2026.
Table of contents
- Key takeaways
- How Deep AI Now Runs Inside Manufacturing Work
- Why Shadow AI Is Falling but Account Switching Has Nearly Tripled
- Regulated Data and AI Agents Are the New Exposure
- How Attackers Turned AI Answers Into a Delivery Route
- What Netskope Recommends, and Why to Read It Critically
- Is manufacturing getting AI risk under control?
- What to be careful about
- Frequently asked questions
How Deep AI Now Runs Inside Manufacturing Work
Adoption across manufacturing is broad and concentrated in a handful of platforms. Microsoft Copilot is the most widely used AI application in the sector, present at 90% of organisations, with Claude Platform close behind at 89% and ChatGPT at 76%. Netskope says Copilot's lead sets manufacturing apart from other industries, though it gives no comparison figures for those sectors.
The larger exposure comes from AI that nobody deliberately opens. Netskope found AI features embedded in software used by 96% of employees, and 90% of staff touch systems that train models on customer or user data. Only 64% use AI applications directly. Much of the sector's AI exposure sits inside ordinary software, running in the background where it is hard to see what information goes in.
Netskope's own numbers do not line up exactly. It puts active AI users at 67% of the sector in one place and direct AI application users at 64% in another, without explaining the gap. Both readings point the same way: roughly two in three manufacturing workers now use an AI tool on purpose, and nearly all of them meet one by default.
Why Shadow AI Is Falling but Account Switching Has Nearly Tripled
Shadow AI means staff using AI tools through personal accounts that the employer cannot see or govern. Manufacturers have pushed hard against it. Personal AI app use dropped from 75% to 55% over the year, while managed AI use rose from 35% to 70%.
The weak spot is movement between the two. The share of users hopping between personal and enterprise accounts went from 9% to 25%, close to three times the earlier level. Netskope reads this as employees trying new tools faster than companies approve them. It argues that quicker review of new AI apps, paired with tight instance-level controls, would reduce the switching, and that because shadow AI seldom goes away completely, firms need durable guardrails, clear policies and an easy way to request tools.
Blocking is the other lever. Particular Audience is the AI app manufacturers restrict most, at 45% of organisations, followed by Emergent at 39% and DeepSeek at 36%. Netskope links those blocks to concerns about where prompts and uploads end up, which internal systems an app can touch, and whether processing runs outside existing security and compliance controls.
Regulated Data and AI Agents Are the New Exposure
When manufacturing staff trip AI-related data policies, regulated data is the most common type, at 57%. Passwords and keys follow at 19%, intellectual property at 16% and source code at 9%. The ranking matters because the most frequently exposed category is also the one that carries legal and compliance consequences, and leaked keys can open systems directly.
Agent traffic is the fastest-growing piece of the picture. The count of non-human entities connecting to remote MCP servers rose 490%, and MCP events grew 600%. In Netskope's framing, MCP servers are connection points through which AI agents reach other users, systems and data.
Netskope's conclusion is that inventory is no longer enough. Security teams used to ask which AI apps and agents were running. They now also have to know who and what those agents communicate with and what they can open. That is a harder question, because an agent's reach depends on the permissions behind every connection it makes, not on the tool a person installed.
How Attackers Turned AI Answers Into a Delivery Route
Netskope describes a malicious AI link as a dangerous URL handed back by an AI tool, which a person then clicks or an agent then follows. It compares the tactic to poisoned search results, which attackers push through SEO, paid advertising or hijacked legitimate websites.
In manufacturing, those interactions began rising in July 2025 and reached 150 per 100,000 users in October 2025. A second climb started in February 2026 and hit 200 per 100,000 in March. AI-themed lures aimed at corporate users moved in a similar rhythm, spiking in February 2026 and rising again in June 2026.
The method is evolving. Netskope points to artificial intelligence engine optimization (AIEO), which it describes as attackers trying to influence what content AI models surface, and notes that advertising is starting to appear inside AI applications. Its practical answer is to treat links from AI tools with the same suspicion as links in email or search results, checking them before anyone follows them.
What Netskope Recommends, and Why to Read It Critically
Netskope's first priority is inspecting every HTTP and HTTPS download across web and cloud traffic. Next, it says, block apps with no business purpose or outsized risk, beginning with a policy that permits reputable apps already in use and blocks everything else. Data loss prevention (DLP), software that spots sensitive content such as source code, regulated data, passwords and keys before it leaves for personal or AI apps, should be central. Remote Browser Isolation (RBI) adds a layer for visits to newly observed or newly registered domains.
Netskope also urges visibility over managed and unmanaged AI apps and API calls, real-time guardrails on prompts and data, tracking of adoption and data loss incidents, and red teaming of private AI deployments before they reach production.
Two caveats apply. Netskope sells products named in the advice, including Netskope One NG-SWG, which it says can apply threat protection to downloads in every category and file type. And every figure comes from its own customers, not a random sample of manufacturers, so the percentages describe firms that already run a Netskope gateway. No independent researcher has been cited assessing the findings.
| Measure | A year earlier | Latest |
|---|---|---|
| Active AI users | 40% | 67% |
| Organisation-managed AI app use | 35% | 70% |
| Personal AI app use | 75% | 55% |
| Users switching personal and enterprise accounts | 9% | 25% |
Is manufacturing getting AI risk under control?
The case for
- Managed AI use doubled from 35% to 70% while personal AI use fell from 75% to 55%, shrinking shadow AI.
- Organisations are actively blocking apps they judge risky, led by Particular Audience at 45%, Emergent at 39% and DeepSeek at 36%.
- DLP and download inspection target the exact categories showing up in violations, including regulated data, keys and source code.
The case against
- Account switching between personal and enterprise AI rose from 9% to 25%, so data still moves through ungoverned accounts.
- AI features are embedded in software used by 96% of employees, much of it outside direct oversight.
- Non-human entities connecting to remote MCP servers rose 490%, widening the paths agents can use to reach data.
- Malicious AI link interactions climbed to 200 per 100,000 users in March 2026 as attackers explore AIEO and in-app advertising.
What to be careful about
- Regulated data makes up 57% of AI-related policy violations, exposing manufacturers to compliance breaches when it reaches AI tools.
- Passwords and keys account for 19% of violations, and a leaked credential can give direct access to production or business systems.
- Fast growth in agent connections to remote MCP servers means agents may reach systems and data beyond what security teams have mapped.
- Links returned by AI tools can deliver malware or phishing to both staff and agents unless they are inspected before being followed.
The bottom line
The Netskope data describes a sector that has largely moved AI under company control and still lost sight of where it runs. Managed use doubled to 70%, yet 96% of staff meet AI inside other software, account switching reached 25%, and agent connections to remote MCP servers rose 490%. Regulated data leads the violations, and attackers have learned to plant links in AI answers. The most important shift is from tools to connections: approving a chatbot is a familiar task, while tracking everything an agent can reach is not. Readers should weigh that all of this comes from one vendor's customer base, and that the same vendor sells the fixes. Independent data would show whether the pattern holds more widely.
What to watch
- Watch whether Netskope's next manufacturing figures show the 25% account-switching share falling, which would test its argument that faster app approval curbs shadow AI; no date has been set.
- Watch for malicious AI link interactions in manufacturing climbing above 200 per 100,000 users, the peak in Netskope's current data; no date has been announced for its next update.
- Watch for independent data on agent connections to remote MCP servers in manufacturing that could confirm or challenge Netskope's 490% figure; none has been scheduled.
Frequently asked questions
What data is the Netskope manufacturing AI report based on?
Netskope Threat Labs used anonymised telemetry from Netskope One customers in manufacturing, collected between 1 June 2025 and 31 August 2026. Because it covers only Netskope's own customers, it reflects firms already using its security gateway rather than the whole sector.
What is AIEO in cybersecurity?
Netskope describes artificial intelligence engine optimization (AIEO) as an emerging technique in which attackers try to influence the content AI models surface to users. It is the AI-era counterpart of pushing malicious links into search results through SEO, and Netskope ties it to malicious AI link interactions reaching 200 per 100,000 manufacturing users in March 2026.
What does Remote Browser Isolation do for manufacturers using AI?
Netskope recommends Remote Browser Isolation (RBI) as an extra layer of protection when employees visit higher-risk sites, such as newly observed or newly registered domains. It sits alongside download inspection and data loss prevention in Netskope's list of controls.
Related reading