Google Cloud: the Vertex AI safety stack

Lesson 4 of 5 in Guardrails in the Cloud: Implementing Controls on AWS, Azure, and GCP.

Vertex AI is Google Cloud’s model platform, and its first guardrail layer ships inside the model API itself: configurable safety filters on Gemini models. Responses are scored against harm categories — harassment, hate speech, sexually explicit content, and dangerous content — and you set a blocking threshold per category, from block low and above (strictest) through block medium and above to block only high and, for approved cases, off. Two details matter for governance. First, the API returns the safety ratings with every response, so your gateway can log the scores even for traffic it allows — a running record of how close to the line your application lives. Second, some protections are non-configurable: certain categories of egregious content (child sexual abuse material among them) are blocked regardless of your settings — a floor the customer cannot trade away.

For hallucination control, Vertex offers grounding: anchor Gemini’s answers in Google Search results or in your own data (via Vertex AI Search over your document corpus), with responses carrying citations and grounding-support metadata your application can inspect and threshold. Same governance pattern as AWS and Azure — hallucination handled as a measurable, gateable quantity — with the distinctive option of grounding against the live web.

Around the model sit the familiar companions. Vertex AI model monitoring watches deployed models for input drift and training-serving skew, and Google’s generative AI evaluation tooling scores outputs on quality and safety criteria. Audit trails run through Cloud Audit Logs, and IAM handles least-privilege access to models and their configurations.

One piece of history worth knowing in a governance course: the model card itself is a Google invention — the 2019 paper Model Cards for Model Reporting (Mitchell et al.) proposed the standardised model fact sheet that SageMaker Model Cards, Hugging Face model cards, and half the documentation requirements in modern AI law now descend from. When the EU AI Act demands technical documentation and instructions for use, it is asking, in legal language, for a rigorous model card.

One guardrail vocabulary, three clouds (verify names against current docs)
CapabilityAWSAzureGoogle Cloud

Harm-category content filtering

Bedrock Guardrails content filters (6 categories, tunable strength)

Azure AI Content Safety (4 categories, severity 0–6 thresholds)

Vertex AI configurable safety filters (per-category block thresholds)

Injection & jailbreak defense

Prompt-attack filter within content filters

Prompt Shields (direct + indirect/XPIA)

Model Armor prompt screening

Hallucination / grounding check

Contextual grounding checks (grounding + relevance thresholds)

Groundedness detection (with correction mode)

Grounding with Google Search or your data; grounding-support metadata

PII / sensitive-data handling

Sensitive-information filters (block or mask; custom regex)

Purview DLP + PII detection via Azure AI Language

Model Armor with Sensitive Data Protection integration

Model-agnostic policy API

ApplyGuardrail API

Content Safety standalone APIs

Model Armor (model-independent)

Evaluation & red-teaming

Bedrock model evaluation; SageMaker Clarify

Foundry evaluators, adversarial simulator, AI red-teaming agent (PyRIT)

Gen AI evaluation service

Invocation logging & audit

CloudWatch metrics, CloudTrail, model invocation logging

Application Insights tracing (OpenTelemetry), Azure Monitor, Purview audit

Cloud Audit Logs, safety ratings returned per response

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.