Cross-referencing & collaboration: AgentShield ↔ AgentSeal

[elliotllliu]

Hi AgentSeal team! 👋

I'm the maintainer of AgentShield, another open-source security scanner for AI agents and MCP servers. I noticed we have very similar goals — protecting the AI agent ecosystem from malicious skills and tools.

What AgentShield brings

• 30 static analysis rules with AST taint tracking (Python ast module, not regex)
• Cross-file data flow analysis — follows imports to trace secrets → HTTP exfiltration paths
• Kill chain detection — 5-stage attack model (Recon → Access → Collection → Exfil → Persistence)
• Multi-language prompt injection — 55+ patterns in 8 languages
• Description-code integrity — detects tools that lie about their capabilities
• Dify .difypkg support — auto-extracts and scans plugin archives
• CI/CD native — GitHub Action + --fail-under scoring + JSON output

Collaboration ideas

1. Benchmark cross-validation — Would love to test AgentShield against SealBench and vice versa. Shared benchmarks benefit the whole community.

2. Complementary strengths — AgentSeal's shield (real-time monitoring) + scan (prompt injection testing) complement AgentShield's static analysis + cross-file analysis. Perhaps we could cross-reference each other?

3. Shared threat intelligence — As we both discover new attack patterns in the wild, sharing them would help the whole ecosystem.

What do you think? Would love to connect.

🔗 https://github.com/elliotllliu/agent-shield
📦 npx @elliotllliu/agent-shield scan ./path/