Developer builds preflight scanner to catch MCP tool security flaws before deployment

·1 views

A developer has created an open-source preflight security scanner targeting Model Context Protocol (MCP) tools, addressing gaps that standard demos fail to reveal. The tool runs static and behavioral checks across four rule categories, flagging issues such as unsafe shell commands, excessive scope declarations, embedded secrets, and unvalidated user inputs. Behavioral tests go further by calling a local fixture server to verify tenant isolation, write approval enforcement, and quota limits. Each finding is assigned a severity level, a remediation step, and a trackable status, turning security observations into actionable engineering tasks. The scanner is positioned as a bounded first-pass filter rather than a full penetration test, with source code available on GitHub at github.com/glatinone/mcp-security-preflight.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How Node.js Webhook Consumers Should Handle Rate Limits and Dead-Letter Queues

A technical guide for B2B SaaS platforms warns that successful HTTP delivery responses do not confirm that downstream business actions—like renewal reminders—actually completed on time. The core principle is that a consumer must treat the business deadline as the primary service-level objective, making rate limits, retry policies, and dead-letter decisions explicit parts of the system contract. Engineers are advised to define three key numbers before choosing queue technology: the business deadline, the sustainable downstream processing rate, and the maximum tolerable duplicate effect. The guide distinguishes between acknowledgement, negative acknowledgement, and dead-lettering as separate outcomes based on the future likelihood of success, not HTTP status codes. Duplicate handling is flagged as an application-level responsibility under at-least-once delivery, with uniqueness constraints on event IDs recommended to prevent stale or repeated actions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Stale Retry Silently Restored Access to a Deactivated Account

A technical case study highlights how a timed-out provisioning job can inadvertently restore access to a deactivated user account in distributed systems. In the described scenario, access was correctly removed from a deactivated account, but an older queued worker woke up and replayed the original group-addition request minutes later, bypassing the deactivation. Each individual system component behaved as designed, yet the combined sequence produced an incorrect and insecure final state. The article argues that queues preserve work but cannot guarantee that queued work remains valid, making generation-based checks and idempotent reconciliation essential safeguards. It also warns that provisioning systems often treat delayed onboarding as more urgent than delayed removal, leaving access-revocation gaps that are harder to detect through standard API success metrics.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Securely Bridge Webhooks to Kafka, Redpanda, and NATS at the Edge

Directly ingesting webhooks into event brokers like Kafka or NATS exposes systems to synchronous timeout failures, replay attacks, and duplicate deliveries. An edge-gateway pattern addresses this by terminating TLS, validating provider-specific HMAC signatures from services like Stripe, GitHub, and Shopify before any payload parsing occurs. Signature checks must use constant-time comparisons and run against the raw request body, while deduplication relies on provider-supplied delivery IDs. Verified events are then normalized into the CloudEvents standard format and routed to the appropriate broker partition based on entity identifiers. Observability metrics and dead-letter queue handling are also essential components of a production-grade implementation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How MCP Servers Must Change When Moving From Local to Multi-Tenant Deployments

Model Context Protocol (MCP) servers that run locally on a single machine behave very differently once deployed behind a gateway accessible to multiple users or teams. A developer built a local lab environment to surface the key problems that emerge in this transition, including authentication, tenant data isolation, and write-tool approvals. The core findings suggest that bearer token validation must happen at the gateway level on every request, and tenant scoping should be enforced at the edge rather than inside individual tools to prevent data leakage. Write operations require time-bound, request-specific approvals, meaning a stale or mismatched approval cannot authorize unintended actions. The author distilled the lessons into a five-point checklist covering per-request auth, edge-level tenancy, expiring approvals, structured audit logs, and pre-tested failure modes.

0 comments Read more at DEV Community