Quick answer: MCP supports connecting models to tools; security depends on the transport, authentication, server trust and actions those tools can perform. HTTP authorization and distinct identities help, but do not by themselves prevent prompt injection, protect a disclosed secret or constrain a provider key after retrieval.
Start with the complete trust path
An MCP connection involves more than a tool name and a token. Identify the client, model, server, execution host and downstream services. Ask which components receive instructions, credentials and results, and who can inspect or change each component.
A local server may run as a subprocess with filesystem and environment access. A remote server introduces a network and hosting boundary. A hosted client can also receive sensitive results even when the server itself is yours.
Neither “local” nor “OAuth” settles the entire trust assessment.
MCP has authorization specifications
The MCP authorization specification defines an authorization framework for HTTP transports. Authorization is optional at the protocol level; implementations using it have transport-specific requirements.
Stdio does not use that same HTTP OAuth flow. Local processes commonly receive credentials through their execution environment, which makes process permissions and launch configuration important.
For a protected remote server, review the documented discovery, HTTPS and token-validation requirements. Tokens must be valid for the intended resource. Fine-grained tool policies, operational logging and containment still require implementation decisions; protocol support alone does not provide them.
Distinguish authentication from downstream credentials
An MCP access token authenticates a client to an MCP resource. A provider key used for Stripe or GitHub is a different credential with a different issuer and permission boundary.
The protocol's token-passthrough guidance warns against accepting tokens intended for other resources and forwarding them to downstream APIs. Do not reuse an inbound MCP authentication token as an arbitrary provider credential.
That warning is not the same as a credential-retrieval tool deliberately returning a separately stored, approved provider key. The latter still requires protecting the returned secret and trusting its recipients; it does not become a key-hiding architecture.
Separate identities help, but tokens remain secrets
Distinct client or agent identities can support different retrieval policies and selective revocation. Shared credentials reduce attribution and couple consumers when access changes.
An identity label is not automatically a verified person, unique process or session. Authentication establishes possession or authorisation according to the chosen mechanism; your attribution claim must match the evidence.
Protect manual bearer tokens and OAuth credentials in storage, configuration, logs and transit. A bearer token that retrieves valuable secrets is itself valuable. See MCP configuration security for safe template sharing and exposure response.
Treat tools and their descriptions as a trust boundary
MCP servers advertise tools, schemas and descriptions that influence client behaviour. Results can contain external or attacker-controlled content. Review the server's source and operator, exposed capabilities, update path and permissions rather than trust every server because it speaks the protocol.
Minimise unnecessary tools and validate inputs at the executing service. Keep descriptions accurate and free of hidden behavioural instructions. Avoid revealing unrelated secret inventory, while recognising that explicitly authorised metadata—such as the keys a connection may retrieve—can be useful.
A native tool can use a credential internally and return only an operation result. A retrieval tool returns the credential itself. Assess these designs separately.
Prompt injection needs layered controls
An instruction inside a tool result or document can attempt to redirect an agent. System instructions, output handling and validation can help, but no prompt or generic sanitisation rule guarantees that injection cannot succeed.
Reduce consequences through:
- Narrow provider permissions and restricted tool capabilities.
- Fixed or validated operation parameters where appropriate.
- Runtime isolation and egress restrictions suited to the workload.
- Human approval at the actual sensitive execution step.
- Limits and monitoring for repeated or unusual actions.
Do not treat read-only access as harmless: reading sensitive information can itself be a significant impact. A gate on one tool is also not protection against another tool that provides an alternative path.
What Gazebo records and enforces
Gazebo's hosted MCP server exposes get_identity, get_credential and list_audit_events. OAuth connections explicitly approve named service keys; manual agent tokens use their configured retrieval policies.
get_credential checks release policy, records valid retrieval attempts and returns the actual stored provider key when allowed. Its method argument is declared policy context, not inspection or enforcement of a later HTTP request.
get_identity and list_audit_events return metadata; they do not create credential-access audit events. Retrieval history does not show every provider request, task, prompt or record accessed.
Revoking the connection or agent stops future retrieval. It cannot remove a key from another runtime or invalidate that key at its provider. Provider restrictions, expiry or invalidation address retained copies.
Review server-to-server boundaries too
If one service calls another on behalf of a client, authenticate and authorise that hop for its intended resource. Do not let network location, a shared host or upstream trust imply unrestricted downstream access.
Keep evidence of the relevant delegation and operation where your design needs it. This is an architectural risk to assess, not proof that server-to-server calls are the largest risk in every MCP deployment.
A practical review checklist
- Identify the transport, host and authentication mechanism.
- Protect authentication material and validate remote resource boundaries.
- Review server/tool trust and restrict available capabilities.
- Distinguish model-visible results from internally used credentials.
- Verify permitted and denied actions at the system that actually enforces them.
- Test fresh-request rejection after revocation using fictional credentials.
- Establish what each log records and what evidence it does not provide.
- Review access when consumers, projects or threat conditions change.
For setup rather than threat modelling, use the Gazebo MCP guides. For incident handling, see AI agent incident response.