Netskope Manufacturing AI Report: The 83% Figure Measures Policy Violations
Netskope published its manufacturing AI report on November 10, finding that upstream data-policy violations represented 83% of observed AI-related violations. That finding concerns information entering AI systems, not a verified count of security incidents or successful breaches. The Netskope manufacturing AI report raises a narrower, more useful question: does approving an AI application actually control the information employees send into it?
The distinction changes the management response. A policy alert calls for investigation into data, permissions, and enforcement. A confirmed breach requires evidence of unauthorized disclosure or access. Treating those categories as interchangeable can distort both the threat assessment and the apparent effectiveness of existing defenses.
The central tension is between approved applications and approved data flows. Microsoft’s security guidance offers a useful reference: protecting an AI workflow requires controls around information and actions, beyond selecting a trusted application. For manufacturing buyers, the question is whether an approved assistant operates inside boundaries that match the task.
What the Netskope Manufacturing AI Report Actually Measures
The 83% finding describes the composition of observed violations, not the percentage of manufacturers suffering AI breaches.
The underlying manufacturing AI findings come from anonymized Netskope platform telemetry covering a subset of manufacturing customers. The observation period ran from June 1, 2025, through August 31, 2026. The public methodology does not disclose an exact customer count.
That makes this an observational vendor dataset, rather than an industry census or employee survey. Its value lies in showing activity visible through deployed controls. It cannot establish the prevalence of every AI risk across every manufacturer.
An upstream violation concerns data entering an AI workflow. A downstream violation concerns information leaving that workflow through generated material or subsequent use. Neither label, by itself, establishes whether someone successfully stole information.
Consider an illustrative employee who tries to upload a confidential supplier document. A security control detects the document and blocks the transfer. The resulting alert provides evidence that someone attempted a prohibited action, while also indicating that the control intervened.
Compare that with an upload that proceeds and later receives an alert. Investigators must determine what crossed the boundary, which account received it, and what the destination retained. Similar policy labels can conceal materially different outcomes.
This distinction matters when executives ask whether rising alerts mean worsening security. More alerts can reflect more risky activity, broader monitoring, stricter rules, or improved detection. An alert total cannot resolve those possibilities without supporting operational evidence.
The same caution applies to a falling count. Fewer alerts might indicate safer behavior. They might also follow weaker inspection, disabled rules, or a shift toward channels that monitoring does not cover.
For a manufacturing security leader, the useful follow-up is therefore operational. Which transfers were prevented, which were permitted, and which remain unresolved? Those categories reveal more about protection than a dramatic headline percentage.
The report’s publication date also differs from its measurement window. November readers are seeing an analysis of activity ending in August. That timing matters when comparing the findings with a company’s current deployment or recently introduced controls.
The appropriate interpretation is an exposure signal. The dataset directs attention toward the entry point where business information meets AI processing. It does not supply a shortcut for estimating confirmed losses, compromised factories, or the likelihood of production disruption.
Approved AI Leaves Security Teams With a Data Decision
Application approval settles who supplies the tool; it does not settle which business records belong inside each task.
Netskope reports that organization-managed AI use rose from 35% to 70%, while personal-account use declined from 75% to 55%. Those figures describe overlapping usage patterns rather than mutually exclusive groups. They suggest migration toward organizational oversight without establishing that personal usage has disappeared.
For manufacturing AI governance, that creates a practical ownership problem. Procurement can approve a supplier. Technology teams can deploy an application. Someone still needs authority to decide which documents, identifiers, and business processes the deployment should include.
An illustrative procurement assistant shows the gap. Summarizing a public supplier brochure presents a different information decision from summarizing a confidential negotiation history. Both requests can occur inside the same approved interface.
A broad application allowlist treats those requests similarly unless another control distinguishes their contents. A defensible deployment needs rules that connect the destination, account, information category, and purpose. Approval must survive contact with the actual work.
That puts pressure on several groups at once. Security teams need enforceable boundaries. Business owners need usable workflows. Legal and privacy teams need processing terms that correspond to the information employees actually submit.
The immediate challenge is coordination, rather than choosing which group should own every decision. A document owner understands business sensitivity. An application administrator understands configuration. Security staff understand detection and response. Effective decisions require those perspectives to meet.
An approved account also needs a clear identity boundary. Employees should understand whether they are operating inside an organizational workspace or a personal session. A recognizable product logo does not answer that question.
The operational issue extends beyond deliberate uploads. A workflow might collect information through an attached document, a connected repository, or material assembled automatically for a request. Each route deserves an explicit decision about access and destination.
This is why training that says “use approved AI” can leave employees uncertain. The instruction identifies an application but leaves the information judgment unresolved. Workers still need concrete guidance tied to their responsibilities.
For example, a company could distinguish approved public documentation from restricted customer correspondence. The exact categories depend on the business. The governing principle is that employees should not need to interpret an abstract security policy during every routine task.
The Netskope manufacturing AI report therefore places pressure on the definition of deployment success. Assigning accounts is an adoption milestone. Demonstrating that those accounts handle only appropriate information is a separate security milestone.
For enterprise buyers, the commercial implication follows directly. A deployment proposal should explain who owns data decisions, how exceptions work, and what evidence administrators can inspect. A feature list cannot answer those questions.
The Boundary Runs Through Documents, Permissions, and Outputs
The most consequential control point is the information boundary inside the workflow, wherever that boundary occurs.
Netskope AI data security findings invite a shift in evaluation criteria. Instead of asking only whether an assistant is approved, buyers should examine the full path of a representative task. That path starts before the model generates an answer.
An illustrative quality analyst might ask an assistant to summarize a supplier complaint. The source package could contain customer names, contract details, internal commentary, and technical specifications. Only some of that information might be necessary for the summary.
The security decision concerns necessity as well as sensitivity. If the task needs a defect description, it does not automatically need the entire customer correspondence history. Reducing unnecessary context can reduce exposure without preventing the underlying work.
OWASP’s guidance on sensitive information disclosure identifies confidential business information and credentials among the relevant risks. It emphasizes access restrictions and cautions that prompt-based restrictions can be bypassed.
That guidance supplies a useful test for purchasing claims. A vendor saying its assistant follows instructions is describing model behavior. A system preventing access to an unauthorized document is enforcing a separate boundary.
Those mechanisms should complement each other. Instructions can shape an answer, while access controls determine what material becomes available. Neither function should be assumed to replace the other.
Microsoft offers a comparison through its discussion of indirect prompt injection. This attack places malicious instructions inside content an assistant processes, such as a document or webpage.
Microsoft describes layered defenses, including separating untrusted content from instructions and detecting suspicious inputs. Its approach illustrates a broader security argument: a capable assistant still needs protection against the material it reads.
This does not establish that Microsoft, Netskope, or another supplier has solved every relevant problem. It shows different control layers addressing different failure modes. Data inspection and malicious-instruction detection answer related questions, with distinct evidence requirements.
The distinction becomes concrete when an assistant reads an external supplier document. The organization must decide whether the document contains information the assistant may process. It must also prevent embedded text from becoming authority to perform unrelated actions.
Output creates another boundary. A summary might contain information appropriate for an internal investigation but inappropriate for an external supplier response. The fact that the assistant legitimately read a document does not authorize every subsequent disclosure.
Manufacturing AI governance therefore needs a task-level view of permission. Reading, summarizing, sharing, and acting represent different decisions. A deployment becomes easier to assess when the organization can explain each decision independently.
For knowledge workers, the same distinction applies to an AI knowledge base. Making information searchable and deciding who may use it are connected design questions. Better retrieval does not eliminate the need to define an audience.
Agent Connections Make the Boundary Harder to See
An assistant that can retrieve information and invoke tools requires controls beyond the conversation window.
Netskope reports growth of 490% in nonhuman entities connecting to remote MCP servers, alongside 600% growth in related events. These are usage measures, not attack-growth figures. They do not establish that the additional connections were malicious.
Model Context Protocol, or MCP, is a standard for connecting AI applications with tools and data sources. Its relevance here concerns the path between a request and the systems used to fulfill it.
In a simple chat interaction, an employee selects material and submits it. In a connected workflow, an assistant can retrieve additional context through configured tools. The employee’s visible request may describe only part of the eventual information flow.
Consider an illustrative maintenance assistant asked to explain a recurring equipment fault. It might retrieve approved manuals and previous service records. A more extensively connected deployment might also reach purchasing records or external support services.
The business value depends on those connections. So does the permission problem. An assistant should not receive broad access merely because a future request might benefit from additional context.
The protocol’s MCP security practices address implementation risks including authorization failures, session hijacking, and unsafe handling of access tokens. Tokens are credentials that let software access protected services.
These implementation concerns differ from the accuracy of a generated answer. A response can be technically correct while relying on excessive permissions. A useful assistant can still expose an organization through a poorly controlled connection.
For buyers, that changes the meaning of an integration demonstration. Showing that an assistant can retrieve a record proves connectivity. It does not establish that access remains appropriate across different users, tasks, and document categories.
A more informative demonstration would show a legitimate request succeeding and an unauthorized request failing. It would also identify the account used, the permissions applied, and the evidence left for review.
Service ownership matters just as much. A connector without a named owner can remain active after its original business purpose changes. The organization needs a way to understand why access exists and who can withdraw it.
Revocation is especially relevant when a supplier relationship ends or an employee changes responsibilities. A permission model should support removing access without depending on the assistant to remember a written instruction.
The broader manufacturing AI governance issue is visibility across delegated work. When software acts on a person’s behalf, responsibility cannot stop at the person’s initial prompt. It must extend to the systems that execute the request.
The growth figures justify attention to this category, while leaving its absolute scale uncertain. They support asking better questions about connected workflows. They do not justify declaring an equivalent surge in successful attacks.
What the Dataset Cannot Establish
A useful risk signal becomes misleading when readers turn observed alerts into claims about confirmed harm.
The Netskope manufacturing AI report is evidence about activity within its observed environment. It does not establish a causal link between each policy event and an eventual security loss. That gap should remain visible throughout any interpretation.
The absence of a disclosed customer count limits judgments about representativeness. Readers cannot reconstruct the sample’s distribution by organization size, geography, or security maturity from the headline figures alone.
A platform-based dataset also reflects the platform’s visibility. Activity outside monitored routes may receive different treatment or remain unseen. This is a limitation of the measurement approach, rather than evidence that an unseen category is necessarily larger.
Policies introduce another source of variation. A company with strict rules might generate alerts for transfers another company permits. Comparing their totals without comparing their rules risks confusing enforcement choices with underlying exposure.
Alert severity matters as well. A prohibited transfer involving a restricted customer record demands a different investigation from a low-confidence classification match. A category percentage cannot substitute for examining those cases.
Successful prevention also complicates the interpretation. If a system blocks an inappropriate upload, the resulting event can indicate both risky behavior and functioning protection. Calling every such event a breach erases the distinction.
The opposite mistake would be dismissing all alerts as harmless because some were blocked. Repeated attempts can reveal unclear policy, a poorly designed workflow, or persistent pressure to use an unsuitable destination. Prevention does not eliminate the underlying business problem.
Netskope also sells security products relevant to the risks it describes. That commercial context does not invalidate its observations. It does mean buyers should separate evidence about a problem from claims about a particular remedy.
Independent guidance helps maintain that separation. NIST’s generative AI profile places risk considerations across the design, development, use, and evaluation of AI systems. That framing extends beyond any individual product purchase.
The appropriate response is to test controls against a company’s own tasks and information. A relevant evaluation should include permitted activity, prohibited activity, ambiguous content, and operational exceptions. Otherwise, a polished demonstration can conceal unresolved boundaries.
No conclusion about factory downtime follows automatically from the upstream percentage. Establishing operational consequences would require evidence connecting a particular event to affected systems and business impact. The report’s central statistic does not provide that chain.
The strongest reading remains focused: data entering AI workflows deserves scrutiny. The weaker reading turns a policy-observation dataset into an industry-wide incident tally. Buyers gain more from preserving the narrower claim.
What Buyers Should Watch Through Early 2027
The next useful evidence will show whether organizations can expand AI use while making data access more accountable.
The first signal is better reporting on enforcement outcomes. Buyers should look for distinctions between blocked transfers, permitted transfers, investigated cases, and confirmed disclosures. Those categories connect monitoring activity to decisions and consequences.
Evidence of prevented transfers alongside documented resolution would strengthen the case for effective controls. A larger alert dashboard without outcome information would leave that case unresolved. The useful development is clearer accounting, rather than a preferred direction for the raw count.
Organizations can demand similar clarity from internal reporting. An executive summary should identify what the monitoring system observed and what investigators subsequently established. Combining those stages into a single incident number makes the organization harder to assess.
The second signal is whether managed deployments reduce inappropriate movement between account types. Buyers should look for evidence that approved workflows meet employee needs while maintaining the required boundaries.
This does not mean every personal account interaction represents a violation. The important question is whether restricted business information reaches an inappropriate destination. Account context matters because the same application can support different contractual and administrative arrangements.
A deployment that offers a usable, authorized route would strengthen the argument that governance can support adoption. Persistent workarounds would weaken claims that assigning enterprise accounts has resolved the issue.
The third signal is stronger accountability for connected agents and tools. Buyers should look for named owners, limited permissions, reviewable activity, and practical revocation. Product announcements become more meaningful when those capabilities appear in working deployments.
A connector inventory alone would provide an incomplete answer. The organization also needs to know which tasks justify access and whether permissions remain appropriate. Evidence that access changes with responsibilities would strengthen the governance case.
Conversely, broad standing permissions would leave the central problem intact. An assistant might become more useful while the organization loses clarity about its reach. Adoption and control should therefore appear as separate measures in deployment reviews.
For developers, these signals make authorization and event records part of product quality. For business leaders, they make workflow design part of security oversight. For employees, they should translate into clearer decisions about everyday information use.
Netskope AI data security observations are most useful when they lead to those verifiable outcomes. They are less useful as a headline that collapses attempts, alerts, prevention, and confirmed harm into a single alarming category.
Before expanding an assistant’s access, bring a representative business task to the deployment review. Identify the information it needs, the destination it uses, and the action it should perform. Ask the system owner to demonstrate the permitted path and explain where an inappropriate request stops. Then inspect the records that support that explanation. The Netskope manufacturing AI report gives buyers a reason to demand that evidence, rather than treating application approval as the end of governance. The relevant decision is whether an organization can explain and control the movement of its own information. Use that standard when reviewing the next AI rollout.



