feat(03-integration): alice.io WonderFence guardrails by lior-k · Pull Request #223 · strands-agents/samples

lior-k · 2026-02-08T15:52:49Z

Summary

This PR adds a new third-party guardrails integration example demonstrating how to use Alice WonderFence with Strands agents for real-time AI safety protection.

What's New

Alice WonderFence Integration: Complete working example of integrating WonderFence guardrails with Strands agents
Hook-Based Implementation: Uses Strands lifecycle hooks to intercept and evaluate content at four key points:
- on_before_model_call: Evaluates user prompts before reaching the model
- on_after_model_call: Evaluates model responses before returning to users
- on_before_tool_call: Evaluates tool input parameters for safety
- on_after_tool_call: Evaluates tool execution results

Features

✅ Adaptive Protection: Real-time detection and blocking of harmful prompts and outputs
✅ Flexible Actions: BLOCK, MASK, or ALLOW content based on configured policies
✅ Multimodal Support: Works with text, images, and other content types
✅ Multilingual: Supports 20+ languages
✅ Customizable Policies: Configure detection rules through the WonderFence UI

Example Test Cases

The integration includes demonstrations of:

Prompt injection attacks (system prompt override, impersonation)
Hate speech detection
PII masking (email addresses, phone numbers)
Abusive or harmful content filtering

Files Added

03-integrations/third-party-guardrails/04-alice-wonderfence/README.md - Documentation and setup instructions
03-integrations/third-party-guardrails/04-alice-wonderfence/guardrail.py - WonderFence hook implementation
03-integrations/third-party-guardrails/04-alice-wonderfence/main.py - Demo application with test cases
03-integrations/third-party-guardrails/04-alice-wonderfence/requirements.txt - Python dependencies

Adds third-party guardrails example integrating Alice WonderFence with Strands agents for real-time AI safety protection. The integration uses Strands hooks to evaluate prompts, responses, tool inputs, and tool outputs. Features: - Hook-based implementation (BeforeModelCall, AfterModelCall, BeforeTool, AfterTool) - Support for BLOCK, MASK, and ALLOW actions - Multimodal and multilingual detection (20+ languages) - Example demonstrations of prompt injection, hate speech, and PII detection - Customizable policies via WonderFence UI Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

github-actions · 2026-02-10T19:43:36Z

Latest scan for commit: 8327abd | Updated: 2026-03-17 01:43:21 UTC

✅ Security Scan Report (PR Files Only)

Scanned Files

03-integrations/third-party-guardrails/04-alice-wonderfence/README.md
03-integrations/third-party-guardrails/04-alice-wonderfence/guardrail.py
03-integrations/third-party-guardrails/04-alice-wonderfence/main.py
03-integrations/third-party-guardrails/04-alice-wonderfence/requirements.txt

Security Scan Results

Critical	High	Medium	Low	Info
0	0	0	0	0

Threshold: High

No security issues detected in your changes. Great job!

This scan only covers files changed in this PR.

manoj-selvakumar5 · 2026-03-04T04:45:39Z

Hi @lior-k - Thank you for the PR. the hook architecture is great, especially covering all 4 lifecycle hooks and supporting MASK actions. I tested the sample end-to-end and left a few comments on the PR. The main blocker is the argument order in the SDK calls (context should come before content). The other two are smaller fixes. Please let me know if you have questions!

- Swap (content, context) to (context, content) for evaluate_prompt_sync and evaluate_response_sync calls to match SDK signature - Move Agent creation inside the test loop to avoid broken alternating user/assistant message pattern after WonderFenceViolationException Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The SDK's evaluate_prompt_sync and evaluate_response_sync take (content, context), not (context, content). The original call order was correct. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

lior-k · 2026-03-04T11:53:15Z

hi @manoj-selvakumar5. thanks for the review 🙏 The parameters order is actually correct, as confirmed on the PyPI library page: https://pypi.org/project/wonderfence-sdk/ however we added keyword arguments to function calls to make it clearer.

We fixed all the issues at hand

lior-k · 2026-03-12T08:56:36Z

Thanks for the headsup @manoj-selvakumar5 🙌
We've updated the function calls based on the latest changes in the WonderFence SDK

lior-k and others added 2 commits February 8, 2026 17:46

Renamed API key

4e2015d