Integrating Runtime Security

Runnable notebooks showing how to integrate HiddenLayer AI Runtime Security into an agentic system, one per supported payload format, using the HiddenLayer Python SDK.

Notebook	Payload format
evaluate_interaction_openai_chat_completions.ipynb	OpenAI Chat Completions
evaluate_interaction_openai_responses.ipynb	OpenAI Responses
evaluate_interaction_anthropic_messages.ipynb	Anthropic Messages

For the concept, see runtime-security-agent-signals-walkthrough.md: a data-flow diagram and walkthrough of the four boundaries, the per-message signals, and the self-correction pattern.

Quick start

python3 -m venv .venv && source .venv/bin/activate
pip install hiddenlayer-sdk python-dotenv jupyter

cat > .env <<'ENV'
HIDDENLAYER_CLIENT_ID=your-client-id
HIDDENLAYER_CLIENT_SECRET=your-client-secret
ENV

jupyter lab

Get credentials from the HiddenLayer console (Settings, API Keys). Each notebook runs top to bottom with no LLM provider configured.

Beta endpoint; the SDK emits a BetaWarning.

How it works

Call client.runtime.evaluate_interaction() at each boundary where content enters the model's context window (user prompt, tool call, tool result, final answer). The arguments, all recommended for proper functionality:

• interaction: the native provider payload you send to or receive from the model
• metadata: model, provider, requester_id, external_session_id
• hl_project_id: the project whose policy evaluates the interaction
• HL-Runtime-Session-Id header: the same value across the run, so HiddenLayer groups the turns into one session

Each returned message carries analysis.signals (prompt_injection, personally_identifiable_information, code, denial_of_service, guardrails, url, language). Use the signals to decide what your agent does.

Integrating into your agent

The notebook payloads are hardcoded so they run without an LLM provider. In your agent you already build these payloads to call the model; to integrate, pass those same payloads to evaluate_interaction at the matching boundaries. The SDK call does not change, and the agent framework is your choice.

The notebooks show two ways to act on a detection:

• Self-correction: when a signal fires on untrusted input, withhold the flagged content and forward a short security notice built from the signals, so the model self-corrects without ever seeing it and the agent keeps running.
• Policy enforcement: craft HiddenLayer policy rules against these signals; when a rule matches, the decision comes back on outcome, so enforcement happens in the platform rather than your agent code.

Resources

• HiddenLayer Documentation
• Python SDK
• SDK API Reference