model-serving-security

Enforce server-side parameter bounds, rate limits, and SSRF-safe URL handling on model-serving endpoints.

4|Updated Apr 27, 2026
One-click install
npx skills add https://github.com/maruakshay/mii-ai-security --skill model-serving-security
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: model-serving-security
Source: https://github.com/maruakshay/mii-ai-security/tree/main/skills/model-serving-security
Command: npx skills add https://github.com/maruakshay/mii-ai-security --skill model-serving-security

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Model serving endpoints are expensive to operate. A single unthrottled client can exhaust GPU capacity for all users by submitting high-token-count requests. Streaming responses introduce new timing and partial-response leakage channels. Model-generated outputs that include URLs can cause the serving layer to make outbound requests — a classic SSRF vector in a non-obvious location.

Core Features & Use Cases

  • Parameter bounding: Enforce server-side caps on max_tokens, n, logprobs, and streaming to prevent resource exhaustion.
  • Multi-layer rate limiting: Apply per-key, per-user, per-IP, per-organization limits with shared state to deter abuse.
  • SSRF protections: Validate and restrict URLs derived from model outputs before outbound requests; disable URL following by default unless explicitly enabled.
  • Monitoring & quick win: Real-time latency, token usage, and error-rate monitoring with circuit-breaker behavior to maintain availability.

Quick Start

Implement parameter caps, multi-dimensional rate limiting, and SSRF-safe URL handling on your model-serving endpoints today.

Frequently Asked Questions about model-serving-security

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I prevent token-exhaustion DoS on model-serving endpoints?

Prevent token-exhaustion DoS by enforcing server-side parameter bounding on max_tokens, n, and logprobs. This caps request costs and stops a single unthrottled client from exhausting GPU capacity for all users.

How does SSRF protection work for URLs generated in model outputs?

SSRF protection validates and restricts URLs derived from model outputs before any outbound requests occur. It disables URL following by default unless explicitly enabled, preventing unintended server-side requests.

What's the best way to implement multi-layer rate limiting for inference APIs?

Implement multi-layer rate limiting by applying per-key, per-user, per-IP, and per-organization limits with shared state. This deters abuse and prevents multi-dimensional rate-limiting bypass across your endpoints.

Do I need streaming limits to secure model-serving endpoints?

Yes, you need streaming limits to secure model-serving endpoints. Streaming responses introduce timing and partial-response leakage channels, so enforcing server-side bounds on streaming prevents resource exhaustion and data leakage.

Can I use circuit breakers to maintain availability during inference abuse?

Yes, you can maintain availability during inference abuse by using real-time monitoring of latency, token usage, and error rates. Circuit-breaker behavior triggers automatically to protect endpoint availability under heavy load.

Why does my model-serving endpoint make unintended outbound requests?

Unintended outbound requests occur when model-generated outputs include URLs that the serving layer follows. Enable SSRF-safe URL validation to restrict outbound requests before execution and disable URL following by default.