AI-Powered Scans

The HoundDog.ai code scanner can run an AI-powered review over the risky dataflows it detects during a scan. Each verdict can adjust the severity, mark the dataflow as a false positive, and attach a privacy analysis. The review runs before all outputs, the scan results upload, and the fail-severity CI gate, so every report and the Cloud Platform reflect the verdicts.

The review is configured entirely through HOUNDDOG_AI_* environment variables (or the configuration file) and is enabled whenever a provider, model ID, and credentials are set. Disable it for a single run with --no-ai-review.

Note

When the review is enabled, the scanner sends the relevant finding, trace, and source context directly to the AI provider you configure, under your own API key and terms with that provider. Nothing is sent to any AI provider otherwise. Organizations on the Cloud Platform can alternatively run reviews server-side — see AI Integration.

Configuration

Environment Variable	Description
HOUNDDOG_AI_PROVIDER	One of anthropic, aws-bedrock, google-gemini, microsoft-foundry, openai.
HOUNDDOG_AI_MODEL_ID	Provider-specific model ID (for microsoft-foundry, the deployment name).
HOUNDDOG_AI_API_KEY	API key. Required for anthropic, openai, google-gemini, and microsoft-foundry; optional for aws-bedrock (Bedrock API key).
HOUNDDOG_AI_BASE_URL	Base URL override. Required for microsoft-foundry; optional for anthropic, openai, and google-gemini (for example, proxies); ignored for aws-bedrock.
HOUNDDOG_AI_AWS_REGION	AWS region for aws-bedrock. Defaults to the standard AWS region resolution chain.

The API keys below are fake placeholders; replace them with real credentials. The model IDs are examples — use any current model from your provider.

Anthropic

bash

export HOUNDDOG_AI_PROVIDER=anthropic
export HOUNDDOG_AI_MODEL_ID=claude-opus-5
export HOUNDDOG_AI_API_KEY=sk-ant-api03-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
hounddog scan

Other current models: claude-sonnet-5 (faster and cheaper), claude-haiku-4-5 (fastest).

OpenAI

bash

export HOUNDDOG_AI_PROVIDER=openai
export HOUNDDOG_AI_MODEL_ID=gpt-5.6-sol
export HOUNDDOG_AI_API_KEY=sk-proj-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
hounddog scan

Other current models: gpt-5.6-terra (balanced), gpt-5.6-luna (high-volume).

Google Gemini

Uses a Google AI Studio API key:

bash

export HOUNDDOG_AI_PROVIDER=google-gemini
export HOUNDDOG_AI_MODEL_ID=gemini-3.6-flash
export HOUNDDOG_AI_API_KEY=AIzaSyXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
hounddog scan

AWS Bedrock

Authenticates with a Bedrock API key or, when HOUNDDOG_AI_API_KEY is not set, the standard AWS credential chain (environment variables, profiles, IMDS). Model IDs are Bedrock inference profiles:

bash

export HOUNDDOG_AI_PROVIDER=aws-bedrock
export HOUNDDOG_AI_MODEL_ID=global.anthropic.claude-opus-5
export HOUNDDOG_AI_API_KEY=ABSKXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
export HOUNDDOG_AI_AWS_REGION=us-east-1
hounddog scan

Microsoft Foundry

HOUNDDOG_AI_BASE_URL must point to the Foundry resource origin (for example https://<resource>.services.ai.azure.com or https://<resource>.openai.azure.com, without any project path), and HOUNDDOG_AI_MODEL_ID is the deployment name (commonly the same as the deployed model):

bash

export HOUNDDOG_AI_PROVIDER=microsoft-foundry
export HOUNDDOG_AI_MODEL_ID=gpt-5.6-sol
export HOUNDDOG_AI_BASE_URL=https://my-resource.services.ai.azure.com
export HOUNDDOG_AI_API_KEY=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
hounddog scan

Behavior

• The review runs after detection and before all outputs, the scan results upload, and the fail-severity CI gate, so reports and the platform always reflect the verdicts.
• The review never fails a scan. The first dataflow is reviewed as a canary: on a systemic failure (bad API key, unknown model, unreachable endpoint), the remaining reviews are skipped and the scan continues without AI review.
• On cloud scans, verdicts are uploaded with the scan results, and repeat scans of the same branch skip dataflows that were already reviewed. Dataflows without a scanner verdict fall back to the organization’s server-side AI review, if configured.
• Verdicts are also stored in the local scan state, so repeated local scans do not re-review unchanged dataflows.
• Use --no-ai-review to disable the review for a single run without unsetting the environment variables.