Resources are read-only data endpoints exposed through MCP for retrieving context from systems such as databases, knowledge bases, or operational platforms. They separate data access from action execution, which helps preserve auditability and reduce unintended side effects in AI-driven workflows.
What Resources Do in MCP
Resources are the read-only side of MCP: they expose context for retrieval without granting the model authority to change the underlying system. That separation matters because the same workflow can safely consult a database record, document, or operational signal while keeping write actions, side effects, and higher-risk operations outside the resource path.
In practice, this design helps organisations treat context access as a governed interface rather than an informal data scrape. It also makes it easier to reason about what the model was allowed to see, which is important when an AI workflow needs evidence, provenance, or auditability before it takes any action.
How Resources Differ From Tools and Prompts
Resources are not the same as tools. Tools are action-oriented and can create change, while resources are intended to surface information only. That distinction reduces the chance that a retrieval step accidentally becomes an execution step, which is one of the main safety benefits of the MCP model.
Resources also differ from prompt text. Prompt content is often transient and user-supplied, whereas a resource is a structured endpoint with a defined source of truth. When the context comes from a resource, the workflow can anchor decisions in a controlled retrieval path instead of stitching together ad hoc text from multiple places.
This separation is especially useful when an application needs to blend live operational data with natural-language reasoning. The model can inspect the source, but it should not be able to mutate that source through the same interface.
Where Resources Fit in an MCP Workflow
In a typical MCP setup, a client discovers available resources, requests the context it needs, and then uses that context alongside prompts or tool calls. The resource layer therefore sits between raw system data and the model’s reasoning process, providing a clear boundary for read access.
That boundary is valuable when the underlying systems are varied. A single MCP server may expose documents, tickets, configuration records, or telemetry, yet each resource can still be treated as a controlled, read-only source with its own naming, scope, and authorization expectations.
Well-designed resource exposure also supports better traceability. When a model cites or summarises information, teams can identify which resource supplied the context and distinguish retrieval from subsequent action.
Security and Governance Implications of Resource Exposure
Because resources reveal operational context, they can still expose sensitive information even when they are read-only. The main security question is not whether the interface can write, but whether the data behind it is appropriately scoped, classified, and authorised for model consumption.
Resources should therefore be designed with least privilege in mind, especially when they surface business data, configuration details, or environment metadata that could help an attacker map the environment. The exposure risk is often indirect: the data itself is not modified, but it can inform phishing, lateral movement, prompt manipulation, or unsafe decision-making if it is too broad or too detailed.
When resources are used in regulated or high-assurance workflows, teams should also consider audit logging, retention, and source provenance so that downstream decisions can be explained and reviewed.
Risk and Threat Considerations
Read-only resources can still create material security exposure if they reveal secrets-adjacent context, internal topology, or privileged operational detail. The risk is often not data corruption, but information leakage that improves reconnaissance or enables unsafe model outputs.
Failure mechanism: Overbroad resource scopes, weak source authorization, or poor classification can let an AI workflow retrieve more context than it should, including data that helps an attacker understand systems, users, or controls.
Impact: Sensitive context can be exposed through summaries, logs, or downstream prompts, increasing the likelihood of privilege abuse, social engineering, and incorrect automated decisions.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP API Security Top 10 addresses the attack and risk surface, while NIST SP 800-53 Rev 5 and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP API Security Top 10 | API8 — Security Misconfiguration | Resources are exposed endpoints whose scope and access must be constrained. |
| API1 — Broken Object Level Authorization | Resource endpoints can leak data if per-object access is not enforced. | |
| Recommendation — Harden resource endpoints so they expose only approved read-only data and cannot be repurposed for unsafe access. Enforce object-level authorization so each resource returns only data the requester is entitled to see. | ||
| NIST SP 800-53 Rev 5 | AC-6 — Least Privilege | Resource endpoints should limit retrieved context to the minimum needed for the workflow. |
| AU-2 — Event Logging | Resource retrieval benefits from auditability and traceability of context access. | |
| Recommendation — Apply AC-6 to restrict each resource to the minimum data scope required by the AI workflow. Log resource reads so teams can trace which context the model accessed and when. | ||
| NIST Zero Trust (SP 800-207) | SC-7 — Boundary Protection | Resources create a distinct read boundary between data access and action execution. |
| Recommendation — Use boundary protections to separate read-only resource access from higher-risk execution paths. | ||
Practitioner Guidance
Why practitioners should care: Resources are often treated as harmless because they do not execute actions, but the context they expose can still be high value. Treat every resource as a governed read interface with explicit ownership, scope, and review expectations.
What to watch for: The highest-risk situations are broad resources, ambiguous naming, and endpoints that mix operational telemetry with business-sensitive context. Those are the cases most likely to blur the line between safe retrieval and exposure.
Practitioner takeaway: Keep resource endpoints narrowly scoped, review them as part of access governance, and verify that each resource only exposes context the workflow genuinely needs.
Related resources from NHI Mgmt Group
- How should security teams prioritise NHI controls when resources are limited?
- How should security teams govern API clients that manage cluster resources?
- What do teams get wrong about tools versus resources in MCP?
- How should security teams govern authorization when applications add nested resources?
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 4, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org