AI Supply Chain Security: A Practical Risk Guide
AI suppliers can introduce vulnerabilities into your business. Learn what the latest research found and how to assess, monitor, and limit the risk.

Your AI Supply Chain Has a Security Problem. Here's How to Manage It
AI tools can save time, automate routine work, and connect systems that used to sit in separate silos. But once you bring an AI supplier into your business, you're also bringing along its software, integrations, infrastructure, and security practices. If any weak link gets exposed, the impact doesn't have to stop at the vendor - it can spill into your environment.
The risk is easy to miss. A polished demo tells you almost nothing about patching, credential protection, email security, or the permissions an AI tool needs to do its job. New research from Black Kite - based on 18,160 AI-related companies worldwide - finds gaps in several of these basic controls. It's a good reason to tighten supplier oversight, not proof that every AI vendor is compromised.
This guide breaks down what the research does (and doesn't) tell you, how supplier weaknesses turn into real business risk, and how to assess AI tools without turning every procurement into a security engineering project. The goal is a process you can actually run: identify the exposure, focus on the risks that matter, and make a clear decision about access and use.
What is AI supply chain security?
AI supply chain security is the work of identifying and managing the security risks that come with the AI products, services, and components your organization relies on. It's not just about the model. In practice, a business may depend on a vendor's application, cloud infrastructure, APIs, plugins, open-source packages, data providers, and links to internal systems.
The question isn't just, "Is this AI tool secure?" A more useful one is, "What can it reach, who runs the components it depends on, and what happens if one of those pieces is compromised?" This applies whether you're buying a generative AI service, adding an AI feature to an existing product, or letting an agent interact with company systems.
The supplier is part of your operating environment
AI products often need access to business data to be useful. A meeting assistant might process recordings and calendar details. A coding assistant might review source code or work inside a development environment. An agent might read support tickets, update customer records, or call other software via an integration.
Those connections create real relationships between the supplier and your systems. That relationship can involve data transfer, stored credentials, permissions, and dependencies on third-party software. So even if your own network controls are solid, a supplier's security problem can still matter.
Security signals are not the same as proof of compromise
A vulnerability, weak email authentication, or suspicious network activity can point to exposure. But none of those signals - by itself - proves an attacker accessed a supplier's systems or took your data. Risk assessment needs context: severity, whether the issue is exploitable, which assets are affected, and what mitigations are actually in place.
That's especially important when you read broad industry research. Statistics can show a pattern in a category of suppliers. They can't tell you the specific risk of a particular vendor, and they can't predict whether a breach will happen. Use the findings to guide due diligence and prioritization, not to replace evaluating the supplier you plan to use.
What did the research find about AI suppliers?
Black Kite's 2026 AI Supply Chain Report assessed 18,160 companies worldwide that develop, deploy, or support AI-related products and services. The study included organizations across different sizes and activity types - from AI development and implementation to specialized applications and the infrastructure behind AI systems. It excluded companies that use AI only incidentally.
The takeaway isn't "every AI supplier is unsafe." It's that the researchers found weaknesses in controls that matter in everyday business security work: vulnerability management, credentials, email authentication, and network security.
The reported figures worth understanding
According to the report:
- 34.2% of assessed companies had at least one critical, unpatched vulnerability. A critical rating should get attention, but business risk depends on the flaw, whether it's reachable, what systems are exposed, and what mitigations exist.
- 61.7% had DMARC issues, and 37.3% had DKIM weaknesses. These email authentication controls help reduce spoofing and impersonation. When they're missing or misconfigured, phishing that uses a supplier's domain can look more credible.
- 29.7% were rated highly susceptible to ransomware under Black Kite's susceptibility index. That means elevated exposure, not proof that ransomware has hit those organizations.
- 44.3% showed unusual network activity, while 41.8% had botnet-related activity recorded. These are signs worth investigating, not proof that an attacker successfully struck.
- 33.2% had a limited digital footprint. If visibility into domains, subdomains, and IP addresses is thin, external assessment is harder and risk evaluation can miss things.
The report also flagged credential exposure. Stolen corporate logins in infostealer records were linked to 6.8% of the AI suppliers assessed. For companies offering enterprise AI deployment solutions, the figure was 19.9%.
Put the numbers in context
A percentage across a large and varied sample helps spot patterns, but it isn't a score for an individual vendor. The sample includes companies with different products, sizes, customer access models, and exposure levels. A weakness at a supplier that only handles public information isn't the same as a weakness at a supplier with privileged access to production systems.
Use the report as a prompt to ask better questions. Does the supplier have a vulnerability disclosure and patching process? How do they protect customer credentials? What systems can their product access? Those answers tend to matter more than any market-wide statistic on its own.
How can an AI supplier weakness affect your business?
Supplier weaknesses usually become customer impact through access, data, or dependency. A supplier might store customer information, hold credentials, connect to an enterprise platform, or maintain software your organization depends on. If an attacker compromises the supplier, those relationships can become a path toward customer systems or data.
This isn't unique to AI. The difference is that AI tools can be granted useful access quickly - especially when teams experiment outside a formal procurement process. A tool that reads files, searches internal knowledge, or takes actions through an API may create more exposure than a standalone application with no connection to company systems.
Integrations can expand the attack surface
An AI application may depend on a model provider, a hosting platform, external APIs, plugins, and open-source components. Each dependency adds another thing you have to understand and govern. The risk isn't that every dependency is dangerous; it's that the chain can get hard to track as integrations pile up.
Black Kite's analysis of Model Context Protocol components found that 48.8% of 1,732 logged security findings were rated High or Critical. Those findings included code and command injection vulnerabilities. Since MCP can connect AI applications to business systems, the security of the integration layer matters alongside the model and the user interface.
Credentials and permissions determine potential impact
If an AI service has access to a shared drive, customer platform, developer environment, or cloud account, a compromised account - or a vulnerable integration - can have consequences beyond the service itself. The likely impact depends on what the access actually allows. Read-only access to a limited set of documents is different from permission to change records, deploy code, or administer users.
So ask what data the tool can reach, where that data is processed, what the supplier retains, and what actions the tool can perform. Also confirm who owns the credentials and how access gets revoked. These questions turn "vendor risk" into a concrete map of what could be affected.
How should you assess and monitor AI suppliers?
A one-time security questionnaire is a start, not a complete control. Supplier environments shift, product features evolve, and integrations can change after approval. A practical process combines an initial assessment with ongoing monitoring - and a way to respond when something changes.
How much effort you put in should match the supplier's role. A low-risk tool with no sensitive data and no system access doesn't need the same level of review as an AI service integrated with customer records or production infrastructure. Treating every vendor the same wastes time and can turn security into busywork.
Use a risk-based review before adoption
Start by writing down the AI tool, the business owner, the supplier, and why the tool is being used. Then define the boundaries of the relationship:
- Map data: Identify what information the tool receives, including personal data, customer records, source code, and confidential material.
- Map access: List connected systems, account permissions, API scopes, and what actions the tool can take.
- Map dependencies: Ask which model providers, hosting services, integrations, and significant software components support the product.
- Review security practices: Understand how the supplier handles vulnerabilities, patches, credentials, email security, incident response, and customer notification.
- Check operational fit: Confirm that the product's data handling and access model meets your legal, contractual, and internal requirements, including relevant privacy obligations such as GDPR where applicable.
- Record the decision: Note the risks, mitigations, owner, and conditions for continued use.
The supplier's answers should be specific enough to support a decision. If the product's access is unclear, ask for a technical explanation or run a limited pilot with non-sensitive data before expanding use.
Monitor for changes and prioritize findings
Monitoring should match the supplier's importance and access level. For a critical provider, review security information regularly and keep a direct channel for discussing incidents, major product changes, and unresolved findings. For a lower-risk tool, a lighter periodic check may be enough.
When something comes up, prioritize by likely business impact rather than chasing a single severity label. Address actively exploited vulnerabilities and ransomware-related exposures quickly. Then evaluate reachability, whether a fix exists and has been applied, and what data or systems fall within scope.
Keep it practical: assign someone to own each important supplier relationship, define who can approve exceptions, and set a review trigger when the tool's data access or integrations change. That gives security and procurement a way to respond without rerunning a full assessment every time a minor update lands.
When is an AI supplier risk acceptable?
No supplier assessment removes all risk. The real question is whether the tool's value justifies the exposure after reasonable controls are in place. That depends on what the product does, how important it is to the business, what data it handles, and whether a safer alternative can meet the same need.
A useful decision framework is to compare business value, potential impact, and available controls. If a tool saves meaningful time but needs access to sensitive data, you may be able to narrow the scope, use a restricted environment, or run a limited pilot. The point isn't automatically "reject it." If the supplier can't explain data handling - or if permissions can't be constrained - the risk probably doesn't fit that use case.
Reduce exposure before accepting an exception
Start with controls that limit the impact of a compromise:
- Grant the minimum permissions required. Avoid broad administrative or write access when read-only or narrowly scoped access will work.
- Limit data exposure. Keep sensitive datasets out of the pilot, redact what you don't need, or restrict the tool to approved content.
- Separate experimentation from production. Use test accounts and non-production environments until the tool and its integrations are reviewed.
- Control credentials. Use managed authentication where possible, avoid shared secrets, and make revocation straightforward.
- Set a review date or trigger. Reassess if the vendor changes its access model, adds a major integration, or expands the data the tool processes.
These controls don't magically make an untrusted supplier trustworthy. They reduce how much damage a failure could cause - and make the decision easier to revisit.
Make the trade-off explicit
A solid approval record should spell out what the organization is accepting: the intended use, data involved, access granted, known gaps, compensating controls, and the person accountable for the call. That record prevents a temporary pilot from turning into a permanent system with broader permissions than anyone intended.
If the value is modest and the supplier requires extensive access, decline or delay. If the value is substantial and you can contain the risk, proceed with limits and monitoring. The point isn't to eliminate AI supplier risk - it's to avoid accepting it by accident.
Conclusion
AI supplier security is a business and operations issue, not just another vendor questionnaire. AI tools can connect data, people, and systems in ways that create real value - and they can also extend the impact of a supplier's security failure into your environment.
Black Kite's findings support the case for better visibility, but broad statistics should guide questions rather than dictate decisions. Map what each tool can access, assess suppliers based on their role, monitor important relationships, and limit permissions to what the use case requires. When you can explain the benefit, exposure, and controls in plain terms, you can adopt AI deliberately instead of trusting a product demo to answer the security questions.