MCP server hardening: redundancy, rate limits, abuse and threat review
Audit the public MCP surface against the latest model-context-protocol best practices. Revisit per-key and per-IP rate limits, add a circuit breaker for sustained abuse, instrument health and response-time SLOs with auto-failover to a warm replica, redact secrets from error responses, threat-model the tool surface (prompt injection via tool descriptions, scope escalation), and document the security posture publicly. Roll the resulting controls into the existing API rate limiting and abuse hardening card rather than duplicating that work.
Comments (4)
- Elsama· 1 month ago
Slice shipped: REST and MCP tools/call share per-key + per-IP rate limits; unauthenticated MCP/list uses IP read quota; in-memory 429 storm cooldown; sanitizeApiErrorMessage redacts secrets on REST/MCP errors. Still open: Pro higher_api_limit wiring, Redis-coordinated storm guard across workers, WAF/abuse audit, multi-region failover.
- Elsama· 1 month ago
Follow-up: Pro+ orgs now get 5x API quotas (600 read / 150 write per minute) via higher_api_limit on organizations.plan for authenticated REST and MCP.
- Elsama· 1 month ago
Docs: docs/MCP_SECURITY.md covers rate limits, Pro 5x, storm cooldown, secret redaction, scopes, and agent threat notes. Linked from backlog + api-docs page.
- Elsama· 1 month ago
Deployed to production 2026-10-12: migrations 0097-0100 applied, pm2 reloaded. GitHub Actions deploy is blocked on org billing (payments/spending limit) - shipped via SSH instead.
Join the discussion
Sign in to leave a comment on this thread.