Security Skills for AI Agents: Audit, Hardening and Testing

Last updated: 2026-10-29

Security work starts with a specific question: are you reviewing source code, improving a design, evaluating a downloaded Skill, or testing a running service? Each calls for different access and evidence. These five catalog entries all address security tasks, but they do not perform the same kind of review.

The descriptions below reflect the published workflows and their stated requirements. They are not a comparative benchmark, and none of these Skills can establish that a system is free of vulnerabilities.

Choose by the security task

Cloudflare Security Audit: investigate source and trust boundaries

Cloudflare Security Audit is a source-first workflow for security guidance and vulnerability review across codebases, APIs, services and other software. Its entrypoint separates a focused guidance mode from a complete audit with structured artifacts.

Choose it when you need a finding tied to a concrete trust boundary, affected resource and supporting source evidence. A complete audit has a broader process and output requirements; simply asking a security question does not start that process. The instructions require a tightly controlled sandbox for executing target-controlled code, and unresolved cases remain marked as needing validation if those controls are unavailable.

Security and Hardening: build controls into the implementation

Security and Hardening is a task from Addy Osmani's Agent Skills collection. Its instructions cover threat modeling and implementation areas such as authentication, authorization, input handling, dependencies, secrets and AI features.

Use it when designing or changing security-sensitive behavior. Give the agent the intended operation, trust boundaries, supported environment and relevant constraints. The shared references are part of the upstream workflow, so preserve them when installing it. Its checklist helps structure a review; a checked box is not independent proof that a control works.

NVIDIA SkillSpector: review an agent Skill before installation

NVIDIA SkillSpector combines a static scan with a source-aware review of the target Skill's purpose, permissions, scripts and external communication. The Skill Inspector instructions call for an APPROVE, CAUTION or REJECT assessment supported by evidence, rather than a verdict based on the numeric score alone.

The scanner is optional for the workflow: if its CLI is unavailable, the instructions require the agent to say so and continue with a manual review. The default Skill workflow uses --no-llm; optional model analysis needs consent and provider configuration. Dependency names and versions may still be sent to OSV.dev for vulnerability lookups. The review does not execute the target Skill's scripts and cannot guarantee safety.

Web App Penetration Testing: test an authorized running application

Web App Penetration Testing uses Strix for a running web application, with optional source access. Its workflow asks for authorization, an environment, exclusions and test accounts before running active tests.

This is appropriate for an application you own or are explicitly authorized to test. Prefer a staging environment: the upstream instructions warn that active testing can send exploit payloads and create or modify data. Strix needs its own setup and model configuration, or its managed service. Review the generated report, verify findings and check whether a run ended early before drawing conclusions.

API Security Testing: examine endpoints and authorization

API Security Testing is another Strix task, focused on REST, GraphQL and gRPC APIs. Its instructions use an API schema or endpoint information, test identities and explicit exclusions to investigate issues such as object-level authorization and input handling.

Use a test API and accounts representing the roles or tenants you are allowed to exercise. The workflow can take an OpenAPI or Postman specification as a target; it describes a separate path for supplying gRPC definitions. It recommends reviewing the report and reproducing findings before reporting them. The general Strix setup is documented in a separate companion Skill, so installing this task alone does not configure the scanner.

Give the agent a useful scope

Before starting, identify the target and the result you want. For source review, specify the revision and relevant paths. For active tests, identify the authorized environment, permitted accounts, excluded actions and limits. Ask for evidence, affected conditions and an explicit account of what was and was not checked.

A source audit, a scanner report and a penetration test provide different evidence. Use the right workflow for the question, assess findings in context, and verify a fix against the same failure condition. For general correctness review of changes, see Code Review Skills.