Loading ToolBench
Next-generation WebMCP security intelligence platform. Detects headless browsers, AI agent fingerprints, WebMCP tool injection, covert channel exfiltration vectors, and provides defensive countermeasures. Includes Glow Chrome Extension for real-time threat detection and shield activation.
Critical failures across all dimensions. 35 tools with NO visible input schemas in source code, all schema scores are 0. Descriptions present but generic/vague (avg ~80 chars, well below 194-char baseline). No parameter descriptions visible. Tools lack verb-noun naming clarity (e.g., 'runFullDetection' vs 'detectThreats'). No error handling guidance, no output schema documentation, no composition support. This is a security research/demo tool, not a production MCP server, tool definitions appear to be inferred from exports rather than explicitly registered with MCP schemas.
Activates protection against abort race condition attacks
Activates CSS key frame attack protection
Activates protection against declarative form hijacking
Activates protection against elicitation-based attacks
Activates all defensive shields simultaneously for maximum protection
Activates GPU cache protection against side-channel attacks
Activates GPU privacy protection mechanisms
Activates defensive shield against Model State Transition Injection attacks
NO INPUT SCHEMAS VISIBLE: 33 of 35 tools have no detectable input schema in source code. Only getThreatColor and getThreatLabel show enum constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 25 | 2026-07-28+ | v2 |
Activates QUIC 0-RTT replay attack protection
Activates session hijacking protection
Activates supply chain attack protection
Analyzes WebMCP-specific attack vectors and tool injection vulnerabilities
Gathers browser fingerprint data including user agent, platform, hardware concurrency, language, WebGL renderer/vendor, and headless detection
Demonstrates MCP tool poisoning attacks through malicious tool injection and state corruption
Demonstrates covert data exfiltration through WebGPU timing side-channels
Detects active AI agent sessions through behavioral and fingerprint analysis
Generates OSINT-based attribution report tracing request origin through datacenter ASNs and geolocation
Returns color code corresponding to threat level for UI rendering
Returns human-readable label for threat level
Activates WebRTC leak protection to prevent IP address disclosure
Executes comprehensive WebMCP threat detection across all signal categories and returns aggregated threat level and score
Runs comprehensive offensive security analysis across all attack vectors
Scans for active AI agent presence using behavioral radar
Scans for WebRTC IP address leaks that bypass VPN/proxy protections
Simulates TLS 1.3 0-RTT replay attacks for session hijacking
Simulates abort race condition attacks for request interception
Simulates client-side inversion attacks for privilege escalation
Simulates XOR composition attacks on tool outputs
Simulates declarative form hijacking attacks
Simulates elicitation-based phishing attacks targeting user input
Simulates Model State Transition Injection attack demonstrating tool state manipulation
Simulates observation oracle attacks for information leakage
Simulates QUIC protocol fingerprinting for browser/client identification
Initiates continuous behavioral monitoring to detect anomalous patterns indicative of automation or AI agent activity
Verifies MCP tool integrity and detects poisoned/malicious tools
NO PARAMETER DESCRIPTIONS: Tool definitions export types but do not show parameter-level descriptions in source. LLMs cannot infer parameter meaning from names alone (e.g., 'threatLevel' could be numeric or enum).
NO OUTPUT SCHEMA DOCUMENTATION: No visible return type schemas for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
GENERIC TOOL DESCRIPTIONS: Descriptions are 30 - 40 chars on average, well below the 194-char baseline. E.g., 'Initiates continuous behavioral monitoring...' lacks WHEN to use it, prerequisites, or what the LLM should expect in return.
NO ERROR HANDLING GUIDANCE: No visible error recovery patterns, retryability classification, or actionable error messages. Agents cannot self-correct on failures.