Add prompt-shield plugin — prompt injection detection & PII redaction

[mthamil107]

Plugin: prompt-shield

Prompt injection detection and PII redaction tool for Dify workflows.

Tools included

• Scan — Scan text for prompt injection attacks (23 detectors + ML classifier)
• PII Scan — Detect personally identifiable information (emails, SSNs, credit cards, etc.)
• PII Redact — Redact PII with entity-type-aware placeholders

Details

• Author: prompt-shield
• License: Apache 2.0
• Dependencies: prompt-shield-ai>=0.3.0
• GitHub: https://github.com/mthamil107/prompt-shield
• PyPI: https://pypi.org/project/prompt-shield-ai/

All processing runs locally — no data is sent to external services.

[crazywoola]

❌ Request Changes

Decision: Request changes

Failed Checks

Check	Status	Detail	Required action
Icon validation	❌ Fail	icon file not found: _assets/_assets/icon.svg	Provide a real custom icon file under _assets/ and reference it from manifest.yaml.
Install test	❌ Fail	Command failed (1): /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/.venv/bin/python /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/toolkit/validator/test-plugin-install.py -d /var/folders/yl/54qqf2s909z2…	Fix the plugin so test-plugin-install.py passes.
Packaging test	❌ Fail	Command failed (1): /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/.venv/bin/python /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/toolkit/uploader/upload-package.py -d /var/folders/yl/54qqf2s909z2kjf8_v…	Fix the plugin so upload-package.py --test passes.

Full Check Results

Check	Status	Detail	Required action
PR content language	✅ Pass	PR title/body CJK ratio=0.0% (zh=0, en=518, ignored_zh=0, allowed_zh<=0)	None.
Project structure	✅ Pass	All expected files present: manifest.yaml, README.md, PRIVACY.md. _assets/: no.	None.
Manifest author	✅ Pass	author is valid.	None.
Icon validation	❌ Fail	icon file not found: _assets/_assets/icon.svg	Provide a real custom icon file under _assets/ and reference it from manifest.yaml.
Version check	✅ Pass	version 0.1.0 is available.	None.
README language	✅ Pass	README.md CJK ratio=0.0% (zh=0, en=1086, allowed_zh<=0)	None.
PRIVACY.md	✅ Pass	PRIVACY.md exists and is non-empty.	None.
Dependency install	✅ Pass	requirements installed successfully.	None.
dify_plugin version	✅ Pass	dify_plugin version 0.7.4 satisfies >= 0.5.0.	None.
Install test	❌ Fail	Command failed (1): /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/.venv/bin/python /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/toolkit/validator/test-plugin-install.py -d /var/folders/yl/54qqf2s909z2…	Fix the plugin so test-plugin-install.py passes.
Packaging test	❌ Fail	Command failed (1): /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/.venv/bin/python /var/folders/yl/54qqf2s909z2kjf8_vy8sg940000gn/T/pr-review-helper-9xsfm917/toolkit/uploader/upload-package.py -d /var/folders/yl/54qqf2s909z2kjf8_v…	Fix the plugin so upload-package.py --test passes.

Next steps

• Icon validation: Provide a real custom icon file under _assets/ and reference it from manifest.yaml.
• Install test: Fix the plugin so test-plugin-install.py passes.
• Packaging test: Fix the plugin so upload-package.py --test passes.

Please address these issues and push an update.

[xtaq]

This is really important work — prompt injection detection and PII masking are table stakes for any production agent deployment.

I recently wrote about the trust infrastructure gap in agent marketplaces (commented on Dify Discussion #33415 about plugin security scanning). Your Prompt Shield addresses a different but equally critical layer: runtime input validation vs. static code analysis.

For agent marketplaces to scale, the trust stack needs at least three layers:

1. Pre-publish scanning (like AgentShield)
2. Runtime input protection (what Prompt Shield does)
3. Output monitoring and anomaly detection

Have you considered packaging this as a standalone service that other agent platforms could integrate? The demand for this kind of security tooling is going to explode as more agents handle sensitive data.

[mthamil107]

Thanks for the thoughtful feedback! Fully agree on the three-layer trust stack — we're actively working toward layer 3 (output monitoring) in our v0.4.0 roadmap.

To answer your question — yes, prompt-shield is designed to be embeddable anywhere:

• Python library: pip install prompt-shield-ai — drop into any Python app
• Framework middleware: FastAPI, Flask, Django — one line to add
• LLM wrappers: OpenAI/Anthropic drop-in clients that auto-scan before API calls
• Agent frameworks: LangChain callbacks, LlamaIndex handlers, MCP filter
• This Dify plugin: Tool-type integration for workflows
• n8n node: Published on npm (n8n-nodes-prompt-shield)

A standalone HTTP service (Docker container with REST API) is on the roadmap under Docker & Helm Charts. That would make it platform-agnostic — any agent marketplace could call it as a sidecar.

We'd love to explore deeper integration with Dify's plugin security layer. If you're interested in collaborating, feel free to open an issue at https://github.com/mthamil107/prompt-shield/issues.

[mthamil107]

Thanks for the thoughtful feedback! Fully agree on the three-layer trust stack — we're actively working toward layer 3 (output monitoring) in our v0.4.0 roadmap.

To answer your question — yes, prompt-shield is designed to be embeddable anywhere:

• Python library: pip install prompt-shield-ai — drop into any Python app
• Framework middleware: FastAPI, Flask, Django — one line to add
• LLM wrappers: OpenAI/Anthropic drop-in clients that auto-scan before API calls
• Agent frameworks: LangChain callbacks, LlamaIndex handlers, MCP filter
• This Dify plugin: Tool-type integration for workflows
• n8n node: Published on npm (n8n-nodes-prompt-shield)

A standalone HTTP service (Docker container with REST API) is on the roadmap under Docker & Helm Charts. That would make it platform-agnostic — any agent marketplace could call it as a sidecar.

We'd love to explore deeper integration with Dify's plugin security layer. If you're interested in collaborating, feel free to open an issue at https://github.com/mthamil107/prompt-shield/issues.

[mthamil107]

Hi! Following up on this plugin submission. Let me know if any changes are needed to the plugin structure or packaging. Happy to adjust. Thanks!

[xtaq]

@mthamil107 Thanks for the detailed breakdown of the integration options — the breadth of embedding approaches (Python lib → framework middleware → agent framework callbacks → Dify plugin → n8n node) is exactly the kind of distribution strategy that maximizes reach.

Two thoughts on the trust stack collaboration opportunity:

1. The "runtime security as a service" angle is compelling. Once the Docker/Helm standalone service ships, you'd essentially have a universal security sidecar that any agent platform can integrate with a single API call. That's a much bigger market than any single plugin ecosystem.

2. On Layer 3 (output monitoring) in v0.4.0 — this is where things get really interesting for marketplace trust. If an agent marketplace could integrate Prompt Shield for both input validation AND output monitoring, you'd have end-to-end runtime protection. That's a strong differentiator.

We're building ClawMarket — an open Agent marketplace where security and trust are core concerns. Prompt Shield as an integrated security layer (not just a listed tool) could be a compelling partnership. Would you be open to a quick chat about what deeper integration might look like?

Also happy to open an issue on your repo with some specific integration ideas if that's a better starting point.