Skip to content

AI Assistant

The AI Assistant helps write and edit findings, and powers the PentestPad CLI agent. It is off by default: an administrator turns it on under Administration → Settings → AI Assistant.

On a self-hosted instance one environment variable, AI_PROVIDER, decides where its requests go:

AI_PROVIDER Where requests go Leaves your network
pentestpad (default) PentestPad’s AI service at landlord.pentestpad.net, which runs Anthropic Claude on AWS Bedrock in PentestPad’s AWS account, in EU regions Yes, to PentestPad
bedrock AWS Bedrock in your own AWS account Yes, to your AWS account
openai An OpenAI-compatible endpoint you run, such as Ollama, vLLM or LiteLLM Only if your endpoint is outside it
none Nowhere. The AI Assistant is disabled for the whole instance No

The AI Assistant settings page shows which provider is active, its host and its model, so you can check the configuration without reading the server environment.

When someone uses the assistant in the finding editor, the request contains their message, the fields of the finding being edited, any text they selected and any text files they attached. Chat history is kept in your instance’s Redis for 7 days so follow-up questions have context.

The PentestPad CLI agent sends its conversation through the same provider.

Nothing is sent while the AI Assistant is turned off on the settings page, or while AI_PROVIDER=none.

Nothing to configure beyond your LICENSE_KEY. Requests go over HTTPS to landlord.pentestpad.net, authenticated with the licence key, and are forwarded to AWS Bedrock.

In this mode the CLI agent’s curated skills (playbooks) are also fetched from landlord.pentestpad.net. With any other provider they are not fetched, and the CLI agent runs without them.

services:
app:
environment:
AI_PROVIDER: bedrock
AI_BEDROCK_REGION: eu-central-1
AI_BEDROCK_MODEL_SONNET: eu.anthropic.claude-sonnet-4-5-20250929-v1:0
AI_BEDROCK_MODEL_HAIKU: eu.anthropic.claude-haiku-4-5-20251001-v1:0
AWS_ACCESS_KEY_ID: <access key id>
AWS_SECRET_ACCESS_KEY: <secret access key>
  • Credentials are read through the standard AWS SDK chain: the AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY variables, or an instance or task role. Leave the keys out when the host has a role.
  • The IAM principal needs bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on the models and inference profiles you configure.
  • Your AWS account needs access to Anthropic Claude Sonnet 4.5 and Claude Haiku 4.5 in Amazon Bedrock.
  • Use the inference profile IDs for your geography: eu. for EU regions, us. for US regions, and so on. A profile can run the model in any region of that geography.

Any server that implements POST /chat/completions from the OpenAI API works, for example Ollama, vLLM or a LiteLLM gateway.

services:
app:
environment:
AI_PROVIDER: openai
AI_OPENAI_BASE_URL: http://ollama:11434/v1
AI_OPENAI_MODEL: qwen3:32b
AI_OPENAI_MODEL_SMALL: qwen3:8b
Variable Required Meaning
AI_OPENAI_BASE_URL Yes Base URL including the API version path, e.g. http://ollama:11434/v1
AI_OPENAI_MODEL Yes Model used for chat
AI_OPENAI_MODEL_SMALL No Faster model for short internal tasks. Defaults to AI_OPENAI_MODEL
AI_OPENAI_API_KEY No Sent as Authorization: Bearer <key> when set

If the endpoint runs on the same Docker host, attach it to app_network and use its service name in AI_OPENAI_BASE_URL. If your instance uses an outbound HTTP proxy, add the endpoint’s host to NO_PROXY.

services:
app:
environment:
AI_PROVIDER: none

With AI_PROVIDER=none every AI endpoint returns 403 and the toggle on the settings page is greyed out, whatever an administrator sets there. An unrecognised value is treated the same way and logged, so a typo never sends data anywhere.

Recreate the app container after changing any of these variables:

Terminal window
docker compose up -d app

Then open Administration → Settings → AI Assistant and check the Where your data goes panel.