Loading ToolBench
Local-first prompt-injection gate and active-defense proxy for AI agents. MCP-native security server with deterministic guardrail engine, reputation system, and cryptographic audit ledger.
zn provides 15 tools with security-focused functionality. Most tools have descriptions (13/15), but quality is inconsistent. Input schemas are present for all tools but lack depth: many parameters missing type constraints, enums, or validation guidance. Tool naming is generally clear (analyze, shield_scan, check_tool_call) but some are generic (analyze appears twice with different signatures). Output schemas are not documented. Error handling descriptions are absent. The server targets a specialized domain (prompt injection detection, evidence ledger) but definitions do not guide LLM selection or recovery paths.
Hybrid analysis: local deterministic rules (<10µs) with optional cloud neural inspection
Analyze text for security risks (scoring interface for red-team loop)
Send payload to zn Cloud Gate for deep neural multilingual analysis
Pre-check tool invocation arguments against guardrail rules
Post-check tool output for prompt injection and secret leakage
Deterministic guardrail evaluation of input text against prompt injection rules
Export evidence ledger in JSONL or CSV format with cryptographic verification for compliance audits
Duplicate tool name 'analyze' with different signatures (text vs input_text, different parameters). LLMs cannot disambiguate; will select wrong variant.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does analyze return? verdict? score? details?).
shield_scan has empty input schema ({}). No parameters documented. Unclear what configuration or filtering is possible.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
Generate aggregated statistics from the evidence ledger
Decorator/context manager for protecting LlamaIndex query engines and agents
Append a cryptographically hash-chained evidence record to the audit ledger
Redact sensitive credentials from text (API keys, passwords, tokens, SSNs, etc.)
Resolve zn Cloud Gate endpoint URL from options or environment
Sanitize tool output by redacting secrets and removing injection payloads
Scan for exposed MCP services on local network
Verify cryptographic integrity of the entire evidence ledger chain from genesis
api_key parameter in analyze_cloud and analyze (hybrid) exposed as tool parameter. Credentials must use server-side injection, not parameters. Secrets will leak into logs and traces.
No error handling guidance in any tool description. LLMs cannot recover from failures (e.g., 'If analysis fails, try local_only=true' or 'If ledger verification fails, check file permissions').