Secure AI Services

Check whether the AI services you run — and the AI tooling your developers use — are configured and governed securely, by running cnspec. This is the posture side of AI security; it complements secure-pipeline, which flags insecure LLM code (prompt handling, untrusted model output) as an agent writes it.

When to use

• Auditing a hosted or self-hosted AI service (OpenAI, Anthropic, vLLM, Databricks AI) and its vector store (Weaviate) for insecure configuration.
• Governing AI tooling: enforcing which AI coding agents, IDE extensions, local LLM runtimes, and MCP servers are approved on a host — the agentic supply chain.
• Checking MCP server/client security (tool-schema validation, prompt-injection / PII / malicious-URL content checks).

What cnspec checks

• AI services — mondoo-openai-security, mondoo-anthropic-security, mondoo-vllm-security, mondoo-databricks-security, mondoo-weaviate-security
• AI-tooling governance — mondoo-ai-security (approved agents / IDE extensions / local runtimes), mondoo-mcp-security (MCP server & client security)

These carry compliance/owasp-llm-top-10-2025 mappings, so a scan doubles as an OWASP LLM Top 10 check (strongest on LLM02, LLM03, LLM06).

Run a scan

# Govern AI tooling on the local host (approved agents / MCP servers / runtimes)
cnspec scan local --incognito -f mondoo-ai-security.mql.yaml   -o sarif
cnspec scan local --incognito -f mondoo-mcp-security.mql.yaml  -o sarif

# AI services / vector stores connect via their own provider + credentials —
# load the relevant bundle and follow the cnspec provider docs for the connection.

• Logged in to Mondoo Platform, assigned policies run automatically; otherwise --incognito -f <bundle>.
• Per-service connection and auth are in the cnspec docs — this skill names the right bundle for the outcome; cnspec owns how to reach each service.
• -o sarif for a CI gate; cnspec exits non-zero on failures.

Reference, don't duplicate

cnspec owns the policies and authoring (see the cnspec repo). This skill invokes cnspec and interprets results.

Part of the secure-your-stack toolkit

secure-ai-services is the AI pillar of posture, alongside secure-os (hosts), secure-cloud (cloud accounts), and secure-saas (SaaS & identity). For the code side of LLM security (insecure output handling, secrets in prompts), use secure-pipeline / secure-coding and the LLM Top 10 matrix. Needs cnspec (install) plus access to the target service or host.