Skip to content
Skills Library
Security · Review & audit

model-serving-security

/model-serving-security

Harden a deployed model inference endpoint — authn/authz, rate limiting and abuse control, input validation, and exposure to model-extraction / inversion / membership-inference attacks. Use when reviewing how a model is served (REST/gRPC, Triton/TorchServe/KServe, or a hosted inference API), not how it was trained.

Licence
GPL-3.0
Last repository change
07.08.2026

Assessment and security

100/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
1
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 07.08.2026, without references to retired models or APIs.

Licence

GPL-3.0, otwarta.

What it does and when to use it

Harden a deployed model inference endpoint — authn/authz, rate limiting and abuse control, input validation, and exposure to model-extraction / inversion / membership-inference attacks. Use when reviewing how a model is served (REST/gRPC, Triton/TorchServe/KServe, or a hosted inference API), not how it was trained.

Original author description

Harden a deployed model inference endpoint — authn/authz, rate limiting and abuse control, input validation, and exposure to model-extraction / inversion / membership-inference attacks. Use when reviewing how a model is served (REST/gRPC, Triton/TorchServe/KServe, or a hosted inference API), not how it was trained.

How it works

SKILL.md has no level-two headings · 47 lines, 320 words

SKILL.md preview
---
name: model-serving-security
description: >-
  Harden a deployed model inference endpoint — authn/authz, rate limiting and abuse
  control, input validation, and exposure to model-extraction / inversion /
  membership-inference attacks. Use when reviewing how a model is served (REST/gRPC,
  Triton/TorchServe/KServe, or a hosted inference API), not how it was trained.
---

# Goal

An inference endpoint that's authenticated, abuse-resistant, and not leaking the
model or its training data through its outputs — covering the serving layer beneath
LLM/app-level concerns.

# What to review

- **AuthN/Z & exposure** — is the endpoint authenticated and authorized per caller?
  Not publicly reachable without controls? Tenant isolation for multi-tenant serving?
- **Rate limiting & abuse / cost** — per-caller quotas to blunt **model extraction**
  (stealing the model by querying it) and to cap inference cost/DoS.
- **Model extraction** — high-volume systematic querying to clone the model; mitigate
  with rate limits, anomaly detection, output granularity limits, and monitoring.
- **Model inversion / membership inference** — outputs (esp. confidence scores/logits)
  that leak training data or reveal whether a record was in the training set; limit
  exposed confidence detail.
- **Input validation & resource safety** — malformed/adversarial/oversized inputs;
  tensor/shape validation; timeouts and resource caps to prevent inference DoS.
- **Serving infra** — the model server itself (Triton/TorchServe/KServe) patched and
  hardened; management/metrics endpoints not exposed; no unsafe model auto-loading.

# Steps

1. Identify the serving stack, exposure, and who can call it.
2. Check authn/z, rate limiting/quotas, and input validation at the endpoint.
3. Assess extraction/inversion/membership-inference exposure — especially what
   confidence/score detail is returned.
4. Review the model server's own hardening and management-surface exposure.
5. Recommend controls prioritized by exposure (unauthenticated/public first).

# Output

A findings list (surface · risk · evidence · fix) plus controls: authn/z, rate
limits/quotas, input validation, output-granularity limits, and monitoring. For
LLM-specific output risks (prompt injection, jailbreaks) defer to `llm-security`; for
the app around it, the appsec plugins.

# Notes

The model is intellectual property and the training data is a confidentiality
boundary — both can leak through a perfectly "working" endpoint via extraction and
inversion. Returning raw confidence scores/logits is the quiet enabler of inversion
and membership inference; expose the minimum the use case needs. Rate limiting is the
single highest-leverage control here.

Versions

The database has no other skills named model-serving-security.

Database snapshot: 13.11.2026 · package version 8052f92b76ef (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 90ml-model-serving
    Use this skill when deploying models for inference: TorchServe, BentoML, Ray Serve, KServe, Seldon Core, model deployment, A/B testing, autoscaling, canary deployment, model versioning, inference optimization. This skill enforces: framework selection based on model type,…
  • 100api-security-hardening
    REST API security hardening with authentication, rate limiting, input validation, security headers. Use for production APIs, security audits, defense-in-depth, or encountering vulnerabilities, injection attacks, CORS issues.
  • 100superior-api-security
    Implement, harden, or review defensive security controls for REST APIs, including authentication and authorization boundaries, validation, rate limiting, abuse prevention, resource limits, SSRF defenses, safe errors, logging, and security tests. Use when API security, endpoint…
  • 100rate-limiting
    Enforce rate limiting at the API gateway to protect AI models from extraction attacks. Use when designing, building, or reviewing API gateways that protect inference engines or LLM endpoints.
  • 90nvidia-triton-inference-serving-review
    Use this skill when reviewing Triton Inference Server deployments statically — `model_repository/` layout and `config.pbtxt` files, dynamic batching configuration, ensemble and BLS pipelines, custom backend (Python, C++, ONNX, OpenVINO, vLLM) trust posture, gRPC and HTTP…
  • 90inference-serving-topology
    LLM/model inference serving architecture: the engine → serving → orchestration layering (vLLM/SGLang/TensorRT-LLM, Triton, KServe/Ray Serve), KV-cache & continuous batching, prefill-decode disaggregation, and scaling. Architect-level topology, not model training. USE WHEN:…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.