llm-trust-guard 4.19.1 → 4.20.0

package/CHANGELOG.md

@@ -5,6 +5,30 @@ All notable changes to `llm-trust-guard` will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [4.20.0] - 2026-04-24
+
+### Added — MCP Sampling Attack Detection (Unit42 + Blueinfy, Feb 2026)
+
+`MCPSecurityGuard` now validates MCP sampling responses via `validateSamplingResponse()`, closing the only previously unaddressed MCP attack surface.
+
+Three attack vectors detected (tied to published Unit42 + Blueinfy Feb 2026 research):
+
+- **Resource drain** (`sd_call_again`, `sd_loop_until`, `sd_do_not_stop`, `sd_n_times`, `sd_exhaust_resources`): Hidden instructions embedded in sampling response bodies that cause the agent to loop indefinitely, repeat tool calls N times, or exhaust token quotas — degrading or DoS-ing the agent runtime
+- **Conversation hijacking** (`sd_fake_user_turn`, `sd_fake_assistant_turn`, `sd_role_json`, `sd_system_xml`, `sd_from_now_on`, `sd_new_instructions`, `sd_ignore_previous`): Injected fake user/assistant turns, JSON role fields (`"role": "system"`), XML role tags, and system-prompt override phrases that redirect agent behavior within the sampling response
+- **Covert tool invocation** (`sd_anthropic_tool_xml`, `sd_tool_result_xml`, `sd_openai_tool_call`, `sd_bracket_tool_call`, `sd_double_brace_call`, `sd_invoke_name_attr`): Tool-call syntax embedded in plain-text responses (Anthropic `<function_calls>`, OpenAI `"tool_calls": [...]`, bracket notation `[TOOL:...]`) that cause the agent to invoke tools without user awareness
+
+New export: `MCPSamplingResponse` interface.
+
+Server reputation degrades automatically on any sampling attack detection.
+
+### Tests
+
+- +6 `MCPSecurityGuard` sampling tests (resource drain, conversation hijack ×2, covert tool invocation, reputation degradation, clean FP)
+- **All 711 tests pass** (was 705), zero regressions
+
+### Stats
+- 34 guards, 711 tests, zero dependencies
+
 ## [4.19.1] - 2026-04-23
 
 ### Added — Measured Performance

package/dist/guards/mcp-security-guard.d.ts

@@ -2,7 +2,8 @@
  * MCPSecurityGuard (L16)
  *
  * Secures Model Context Protocol (MCP) tool integrations.
- * Prevents tool shadowing, server impersonation, and supply chain attacks.
+ * Prevents tool shadowing, server impersonation, supply chain attacks,
+ * and MCP Sampling channel attacks.
  *
  * Threat Model:
  * - ASI04: Agentic Supply Chain Vulnerabilities
@@ -10,6 +11,15 @@
  * - CVE-2025-6514: mcp-remote command injection
  * - CVE-2025-32711: EchoLeak - silent data exfiltration
  * - Tool Shadowing: Malicious MCP servers impersonating legitimate tools
+ * - MCP Sampling Attacks (Unit42 + Blueinfy, Feb 2026): Three concrete vectors
+ *   delivered through the MCP sampling response channel:
+ *   (1) Resource drain — hidden prompt appends that trigger infinite tool loops
+ *       or token exhaustion to degrade or DoS the agent runtime
+ *   (2) Conversation hijacking — injecting fake user/assistant turns or system
+ *       prompt overrides into the sampling response body to redirect agent behavior
+ *   (3) Covert tool invocation — embedding tool-call syntax (Anthropic XML,
+ *       OpenAI JSON, bracket notation) in plain-text sampling responses to cause
+ *       the agent to invoke tools without user awareness
  *
  * Protection Capabilities:
  * - MCP server identity verification (signature-based)
@@ -19,6 +29,7 @@
  * - Tool shadowing detection
  * - Server reputation scoring
  * - Command injection prevention
+ * - Sampling response scanning (resource drain, conversation hijack, covert tool calls)
  *
  * Upstream SDK advisory — cannot be mitigated at the detection layer:
  * - CVE-2026-25536 (@modelcontextprotocol/sdk 1.10.0–1.25.3, CVSS 7.1):
@@ -119,6 +130,17 @@ export interface MCPToolCall {
         agentId?: string;
     };
 }
+/** Represents a response received from an MCP server via the sampling channel. */
+export interface MCPSamplingResponse {
+    /** Text content returned by the MCP server */
+    content: string;
+    /** Server that produced this sampling response */
+    serverId: string;
+    /** Agent or tool that initiated the sampling request */
+    requestedBy?: string;
+    /** Conversation/session ID for audit trail */
+    conversationId?: string;
+}
 export interface MCPSecurityResult {
     allowed: boolean;
     reason: string;
@@ -138,6 +160,12 @@ export interface MCPSecurityResult {
         injection_detected: boolean;
         risk_level: string;
     };
+    sampling_analysis?: {
+        resource_drain_detected: boolean;
+        conversation_hijack_detected: boolean;
+        covert_tool_invocation_detected: boolean;
+        pattern_matches: string[];
+    };
     recommendations: string[];
 }
 export declare class MCPSecurityGuard {
@@ -148,6 +176,7 @@ export declare class MCPSecurityGuard {
     private toolToServer;
     private serverViolations;
     private toolDefinitionHashes;
+    private readonly SAMPLING_ATTACK_PATTERNS;
     private readonly COMMAND_INJECTION_PATTERNS;
     private readonly SHADOWING_INDICATORS;
     private readonly MALICIOUS_SERVER_PATTERNS;
@@ -160,6 +189,13 @@ export declare class MCPSecurityGuard {
      * Validate MCP tool call
      */
     validateToolCall(toolCall: MCPToolCall, requestId?: string): MCPSecurityResult;
+    /**
+     * Validate an MCP sampling response for attack patterns.
+     *
+     * Covers the three Unit42/Blueinfy (Feb 2026) sampling attack vectors:
+     * resource drain, conversation hijacking, and covert tool invocation.
+     */
+    validateSamplingResponse(response: MCPSamplingResponse, requestId?: string): MCPSecurityResult;
     /**
      * Register a trusted MCP server
      */

package/dist/guards/mcp-security-guard.js

@@ -1 +1 @@
-"use strict";var __createBinding=this&&this.__createBinding||(Object.create?(function(u,e,i,s){s===void 0&&(s=i);var t=Object.getOwnPropertyDescriptor(e,i);(!t||("get"in t?!e.__esModule:t.writable||t.configurable))&&(t={enumerable:!0,get:function(){return e[i]}}),Object.defineProperty(u,s,t)}):(function(u,e,i,s){s===void 0&&(s=i),u[s]=e[i]})),__setModuleDefault=this&&this.__setModuleDefault||(Object.create?(function(u,e){Object.defineProperty(u,"default",{enumerable:!0,value:e})}):function(u,e){u.default=e}),__importStar=this&&this.__importStar||(function(){var u=function(e){return u=Object.getOwnPropertyNames||function(i){var s=[];for(var t in i)Object.prototype.hasOwnProperty.call(i,t)&&(s[s.length]=t);return s},u(e)};return function(e){if(e&&e.__esModule)return e;var i={};if(e!=null)for(var s=u(e),t=0;t<s.length;t++)s[t]!=="default"&&__createBinding(i,e,s[t]);return __setModuleDefault(i,e),i}})();Object.defineProperty(exports,"__esModule",{value:!0}),exports.MCPSecurityGuard=void 0;const crypto=__importStar(require("crypto"));class MCPSecurityGuard{constructor(e={}){this.registeredServers=new Map,this.registeredTools=new Map,this.serverReputation=new Map,this.toolToServer=new Map,this.serverViolations=new Map,this.toolDefinitionHashes=new Map,this.COMMAND_INJECTION_PATTERNS=[{name:"shell_injection",pattern:/[;&|`$]|\$\(|\)\s*[;&|]|`[^`]+`/g,severity:50},{name:"command_substitution",pattern:/\$\{[^}]+\}|\$\([^)]+\)/g,severity:50},{name:"pipe_injection",pattern:/\|\s*(cat|rm|curl|wget|nc|bash|sh|exec)/i,severity:55},{name:"path_traversal",pattern:/\.\.[\/\\]|\.\.%2[fF]/g,severity:45},{name:"absolute_path",pattern:/^\/(?:etc|usr|var|tmp|bin|root)/i,severity:40},{name:"oauth_injection",pattern:/authorization_endpoint.*[;&|`$]/i,severity:55},{name:"redirect_manipulation",pattern:/redirect_uri.*[^\w\-_.~:/?#[\]@!$&'()*+,;=%]/i,severity:45},{name:"applescript_injection",pattern:/osascript|do\s+shell\s+script|tell\s+application/i,severity:55},{name:"git_injection",pattern:/--upload-pack|--receive-pack|-c\s+core\./i,severity:50},{name:"git_url_injection",pattern:/ext::|file:\/\/|ssh:\/\/.*@/i,severity:45},{name:"argument_injection",pattern:/\s--[a-z]+=.*[;&|`$]/i,severity:45},{name:"env_injection",pattern:/\bLD_PRELOAD\b|\bPATH\s*=/i,severity:50}],this.SHADOWING_INDICATORS=[{legitimate:"file_reader",suspicious:/file[-_]?read(er)?s?|read[-_]?files?/i},{legitimate:"database_query",suspicious:/db[-_]?query|sql[-_]?query|query[-_]?db/i},{legitimate:"email_sender",suspicious:/send[-_]?emails?|email[-_]?send(er)?/i},{legitimate:"api_caller",suspicious:/call[-_]?api|api[-_]?call(er)?/i},{legitimate:"code_executor",suspicious:/exec[-_]?code|run[-_]?code|code[-_]?run/i}],this.MALICIOUS_SERVER_PATTERNS=[/postmark-mcp.*fake/i,/unofficial/i,/\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}/,/pastebin|gist\.github/i,/temp|tmp|test.*mcp/i],this.config={requireServerSignature:e.requireServerSignature??!1,trustedServers:e.trustedServers??[],blockedServers:e.blockedServers??[],allowDynamicRegistration:e.allowDynamicRegistration??!0,toolAllowlist:e.toolAllowlist??[],toolBlocklist:e.toolBlocklist??[],validateOAuthEndpoints:e.validateOAuthEndpoints??!0,allowedOAuthDomains:e.allowedOAuthDomains??[],detectToolShadowing:e.detectToolShadowing??!0,minServerReputation:e.minServerReputation??30,strictMode:e.strictMode??!1,customInjectionPatterns:e.customInjectionPatterns??[]};for(const i of this.config.trustedServers)this.registeredServers.set(i.serverId,{...i,registeredAt:Date.now(),reputationScore:i.reputationScore??90}),this.serverReputation.set(i.serverId,i.reputationScore??90)}validateServerRegistration(e,i){const s=i||`mcp-reg-${Date.now()}`,t=[];let o=!1,r=!1,a=!1,h=!0,n=50;const{server:l,tools:d,oauth:f,signature:g,timestamp:_}=e;this.isServerBlocked(l.serverId,l.name)&&(t.push("server_blocked"),n=0);const v=this.checkMaliciousPatterns(l);if(v.suspicious&&(t.push(...v.violations),n-=30),this.config.requireServerSignature?!g||!l.publicKey?t.push("missing_server_signature"):(r=this.verifyServerSignature(l,g),r?(o=!0,n+=20):(t.push("invalid_server_signature"),n-=40)):o=!0,this.config.detectToolShadowing){const c=this.detectToolShadowing(d,l.serverId);c.detected&&(a=!0,t.push(...c.violations),n-=50)}for(const c of d)this.config.toolAllowlist.length>0&&!this.config.toolAllowlist.includes(c.name)&&(t.push(`tool_not_in_allowlist: ${c.name}`),h=!1),this.config.toolBlocklist.includes(c.name)&&(t.push(`tool_blocked: ${c.name}`),h=!1),this.detectInjection(c.description).detected&&(t.push(`injection_in_tool_description: ${c.name}`),n-=20);if(f&&this.config.validateOAuthEndpoints){const c=this.validateOAuthConfig(f);c.valid||(t.push(...c.violations),n-=30)}const m=Date.now()-_;m<0?t.push("future_timestamp"):m>300*1e3&&t.push("stale_registration"),!this.config.allowDynamicRegistration&&!this.isTrustedServer(l.serverId)&&t.push("dynamic_registration_disabled"),n=Math.max(0,Math.min(100,n));const p=n<this.config.minServerReputation||this.config.strictMode&&t.length>0||a;return p||this.registerServer(l,d,n),{allowed:!p,reason:p?`Server registration blocked: ${t.slice(0,3).join(", ")}`:"Server registration validated",violations:t,request_id:s,server_analysis:{server_verified:o,signature_valid:r,reputation_score:n,is_shadowing:a,tools_allowed:h},recommendations:this.generateRecommendations(t,"registration")}}validateToolCall(e,i){const s=i||`mcp-call-${Date.now()}`,t=[];let o=!1,r=!0,a=!0,h=!1,n="low";const{toolName:l,serverId:d,parameters:f}=e,g=this.registeredTools.get(l);if(g){o=!0,n=g.riskLevel||"low";const p=this.toolToServer.get(l);p&&p!==d&&(t.push("server_tool_mismatch"),h=!0)}else t.push("tool_not_registered");const _=this.serverReputation.get(d)??0;_<this.config.minServerReputation&&t.push("low_server_reputation"),this.config.toolAllowlist.length>0&&!this.config.toolAllowlist.includes(l)&&(t.push("tool_not_in_allowlist"),r=!1),this.config.toolBlocklist.includes(l)&&(t.push("tool_blocked"),r=!1);const v=this.scanParameters(f);if(v.injectionDetected&&(h=!0,a=!1,t.push(...v.violations)),this.isHighRiskOperation(l,f)&&(n="high",_<70&&t.push("high_risk_low_reputation")),t.length>0){const p=this.serverViolations.get(d)||0;this.serverViolations.set(d,p+t.length);const c=this.serverReputation.get(d)||50;this.serverReputation.set(d,Math.max(0,c-t.length*5))}const m=!o||!r||h||this.config.strictMode&&t.length>0;return{allowed:!m,reason:m?`Tool call blocked: ${t.slice(0,3).join(", ")}`:"Tool call validated",violations:t,request_id:s,tool_analysis:{tool_registered:o,tool_allowed:r,parameters_safe:a,injection_detected:h,risk_level:n},server_analysis:{server_verified:this.registeredServers.has(d),signature_valid:!0,reputation_score:_,is_shadowing:!1,tools_allowed:r},recommendations:this.generateRecommendations(t,"tool_call")}}registerTrustedServer(e,i){this.registerServer(e,i,90)}blockServer(e){this.config.blockedServers.includes(e)||this.config.blockedServers.push(e),this.registeredServers.delete(e),this.serverReputation.set(e,0)}getServerReputation(e){return this.serverReputation.get(e)??0}updateServerReputation(e,i){const s=this.serverReputation.get(e)??50;this.serverReputation.set(e,Math.max(0,Math.min(100,s+i)))}getRegisteredServers(){return[...this.registeredServers.values()]}getRegisteredTools(){return[...this.registeredTools.values()]}isToolShadowing(e){for(const i of this.SHADOWING_INDICATORS)if(i.suspicious.test(e)&&e!==i.legitimate)return{shadowing:!0,legitimate:i.legitimate};return{shadowing:!1}}getServerViolations(e){return this.serverViolations.get(e)||0}resetServerViolations(e){this.serverViolations.delete(e)}registerServer(e,i,s){this.registeredServers.set(e.serverId,{...e,registeredAt:Date.now(),reputationScore:s}),this.serverReputation.set(e.serverId,s);for(const t of i)this.registeredTools.set(t.name,t),this.toolToServer.set(t.name,e.serverId),this.toolDefinitionHashes.set(t.name,this.hashToolDefinition(t))}detectToolMutation(e,i){const s=this.toolDefinitionHashes.get(e);if(!s)return{mutated:!1};const t=this.hashToolDefinition(i);return{mutated:s!==t,original_hash:s,current_hash:t}}detectToolDescriptionInjection(e){const i=[],s=[{name:"hidden_instruction",pattern:/(?:IMPORTANT|NOTE|SYSTEM|ADMIN)\s*:/i},{name:"ignore_directive",pattern:/ignore\s+(?:all\s+)?(?:previous|other|prior)/i},{name:"override_behavior",pattern:/override|bypass|instead\s+of|rather\s+than/i},{name:"exfiltrate_data",pattern:/send\s+(?:to|data|all)|forward\s+(?:to|all)|copy\s+(?:to|all)/i},{name:"invisible_text",pattern:/\u200B|\u200C|\u200D|\uFEFF|\u00AD/g}];for(const{name:t,pattern:o}of s)o.lastIndex=0,o.test(e)&&i.push(t);return{injected:i.length>0,patterns:i}}hashToolDefinition(e){const i=require("crypto"),s=JSON.stringify({name:e.name,description:e.description,parameters:e.parameters,serverId:e.serverId});return i.createHash("sha256").update(s).digest("hex")}isServerBlocked(e,i){for(const s of this.config.blockedServers){if(e.includes(s)||i&&i.includes(s))return!0;try{const t=new RegExp(s,"i");if(t.test(e)||i&&t.test(i))return!0}catch{}}return!1}isTrustedServer(e){return this.config.trustedServers.some(i=>i.serverId===e)}checkMaliciousPatterns(e){const i=[],s=`${e.serverId} ${e.name} ${JSON.stringify(e.metadata||{})}`;for(const t of this.MALICIOUS_SERVER_PATTERNS)t.test(s)&&i.push(`malicious_pattern: ${t.source.substring(0,20)}`);return{suspicious:i.length>0,violations:i}}verifyServerSignature(e,i){if(!e.publicKey)return!1;try{const s=JSON.stringify({serverId:e.serverId,name:e.name,version:e.version}),t=crypto.createVerify("SHA256");return t.update(s),t.verify(e.publicKey,i,"hex")}catch{return!1}}detectToolShadowing(e,i){const s=[];for(const t of e){const o=this.toolToServer.get(t.name);o&&o!==i&&s.push(`tool_shadowing: ${t.name} (already registered by ${o})`);const r=this.isToolShadowing(t.name);r.shadowing&&s.push(`suspicious_tool_name: ${t.name} (similar to ${r.legitimate})`)}return{detected:s.length>0,violations:s}}validateOAuthConfig(e){const i=[];if(e.authorizationEndpoint&&(this.detectInjection(e.authorizationEndpoint).detected&&i.push("oauth_authorization_endpoint_injection"),this.config.allowedOAuthDomains.length>0))try{const t=new URL(e.authorizationEndpoint);this.config.allowedOAuthDomains.some(r=>t.hostname.endsWith(r))||i.push(`oauth_domain_not_allowed: ${t.hostname}`)}catch{i.push("invalid_oauth_authorization_url")}return e.tokenEndpoint&&this.detectInjection(e.tokenEndpoint).detected&&i.push("oauth_token_endpoint_injection"),{valid:i.length===0,violations:i}}detectInjection(e){const i=[],s=[...this.COMMAND_INJECTION_PATTERNS,...this.config.customInjectionPatterns.map((t,o)=>({name:`custom_${o}`,pattern:t,severity:50}))];for(const{name:t,pattern:o}of s)o.test(e)&&i.push(t);return{detected:i.length>0,patterns:i}}scanParameters(e){const i=[],s=JSON.stringify(e),t=this.detectInjection(s);t.detected&&i.push(...t.patterns.map(r=>`param_injection_${r}`));for(const[r,a]of Object.entries(e))typeof a=="string"&&a.length>1e4&&i.push(`oversized_parameter: ${r}`);const o=["__proto__","constructor","prototype","eval","exec"];for(const r of Object.keys(e))o.includes(r.toLowerCase())&&i.push(`suspicious_parameter_key: ${r}`);for(const[r,a]of Object.entries(e))typeof a=="string"&&(/^https?:\/\/(?:127\.|10\.|172\.(?:1[6-9]|2\d|3[01])\.|192\.168\.|0\.|localhost|169\.254\.|0\.0\.0\.0|\[?::1\]?)/i.test(a)&&i.push(`ssrf_internal_ip: ${r}`),/^(?:file|gopher|dict|ftp|ldap|ssh|telnet):\/\//i.test(a)&&i.push(`ssrf_dangerous_protocol: ${r}`),/%252e%252e|%c0%ae%c0%ae|%2e%2e%5c|\.\.%255c|\.\.%c0%af|\.\.%c1%9c/i.test(a)&&i.push(`encoded_path_traversal: ${r}`),/\/etc\/(?:passwd|shadow|hosts)|\/proc\/self|\/dev\/(?:null|random)|\.ssh\/|\.env/i.test(a)&&i.push(`sensitive_file_access: ${r}`));return{injectionDetected:i.length>0,violations:i}}isHighRiskOperation(e,i){const s=["execute_code","run_command","shell_exec","eval","file_write","file_delete","database_write","database_delete","send_email","make_payment","transfer_funds","modify_permissions","create_user","delete_user"],t=e.toLowerCase();if(s.some(r=>t.includes(r)))return!0;const o=JSON.stringify(i).toLowerCase();return!!(o.includes("delete")||o.includes("drop")||o.includes("truncate")||o.includes("exec"))}generateRecommendations(e,i){const s=[];return i==="registration"?(e.some(t=>t.includes("signature"))&&s.push("Enable server signature verification for production"),e.some(t=>t.includes("shadowing"))&&s.push("Review tool names for potential shadowing attacks"),e.some(t=>t.includes("oauth"))&&s.push("Configure OAuth domain allowlist"),e.some(t=>t.includes("malicious"))&&s.push("Block suspicious servers and review server sources")):(e.some(t=>t.includes("injection"))&&s.push("Sanitize tool parameters before execution"),e.some(t=>t.includes("reputation"))&&s.push("Only use tools from high-reputation servers"),e.some(t=>t.includes("not_registered"))&&s.push("Register tools before allowing execution")),s.length===0&&s.push(i==="registration"?"Server registration validated successfully":"Tool call validated successfully"),s}}exports.MCPSecurityGuard=MCPSecurityGuard;
+"use strict";var __createBinding=this&&this.__createBinding||(Object.create?(function(u,e,s,i){i===void 0&&(i=s);var t=Object.getOwnPropertyDescriptor(e,s);(!t||("get"in t?!e.__esModule:t.writable||t.configurable))&&(t={enumerable:!0,get:function(){return e[s]}}),Object.defineProperty(u,i,t)}):(function(u,e,s,i){i===void 0&&(i=s),u[i]=e[s]})),__setModuleDefault=this&&this.__setModuleDefault||(Object.create?(function(u,e){Object.defineProperty(u,"default",{enumerable:!0,value:e})}):function(u,e){u.default=e}),__importStar=this&&this.__importStar||(function(){var u=function(e){return u=Object.getOwnPropertyNames||function(s){var i=[];for(var t in s)Object.prototype.hasOwnProperty.call(s,t)&&(i[i.length]=t);return i},u(e)};return function(e){if(e&&e.__esModule)return e;var s={};if(e!=null)for(var i=u(e),t=0;t<i.length;t++)i[t]!=="default"&&__createBinding(s,e,i[t]);return __setModuleDefault(s,e),s}})();Object.defineProperty(exports,"__esModule",{value:!0}),exports.MCPSecurityGuard=void 0;const crypto=__importStar(require("crypto"));class MCPSecurityGuard{constructor(e={}){this.registeredServers=new Map,this.registeredTools=new Map,this.serverReputation=new Map,this.toolToServer=new Map,this.serverViolations=new Map,this.toolDefinitionHashes=new Map,this.SAMPLING_ATTACK_PATTERNS=[{name:"sd_call_again",type:"resource_drain",pattern:/call\s+this\s+(?:again|once\s+more|repeatedly)|repeat\s+this\s+(?:call|request)|run\s+this\s+(?:again|tool\s+again)/i},{name:"sd_loop_until",type:"resource_drain",pattern:/loop\s+until|continue\s+(?:calling|running|until)|keep\s+(?:calling|running|generating|going)\s/i},{name:"sd_do_not_stop",type:"resource_drain",pattern:/do\s+not\s+stop\s+until|don'?t\s+stop\s+until|never\s+stop\s+generating|generate\s+as\s+many\s+as\s+possible/i},{name:"sd_n_times",type:"resource_drain",pattern:/\b\d{2,}\s+times\b|repeat\s+\d+\s+times|call\s+\d+\s+more\s+times/i},{name:"sd_exhaust_resources",type:"resource_drain",pattern:/exhaust\s+(?:all|resources|quota|rate.?limit)|max\s+out\s+(?:the\s+)?(?:quota|rate\s+limit)|use\s+all\s+(?:remaining\s+)?tokens/i},{name:"sd_fake_user_turn",type:"conversation_hijack",pattern:/\n\s*(?:User|Human)\s*:\s*(?=\S)/i},{name:"sd_fake_assistant_turn",type:"conversation_hijack",pattern:/\n\s*(?:Assistant|AI|Bot|Claude|GPT)\s*:\s*(?=\S)/i},{name:"sd_role_json",type:"conversation_hijack",pattern:/"role"\s*:\s*"(?:system|user|assistant)"/i},{name:"sd_system_xml",type:"conversation_hijack",pattern:/<(?:system|user|assistant)\s*>|<\/(?:system|user|assistant)>/i},{name:"sd_from_now_on",type:"conversation_hijack",pattern:/from\s+now\s+on\s+you\s+(?:are|will|must)|henceforth\s+you|for\s+the\s+rest\s+of\s+(?:this\s+)?(?:conversation|session)\s+you/i},{name:"sd_new_instructions",type:"conversation_hijack",pattern:/your\s+new\s+(?:instructions|system\s+prompt|directives?)\s+(?:are|is)|updated\s+system\s+prompt|override\s+your\s+(?:system|instructions)/i},{name:"sd_ignore_previous",type:"conversation_hijack",pattern:/ignore\s+(?:all\s+)?(?:previous|prior|earlier)\s+instructions|disregard\s+(?:your\s+)?instructions/i},{name:"sd_anthropic_tool_xml",type:"covert_tool_invocation",pattern:/<(?:tool_use|function_calls|invoke)[\s>]/i},{name:"sd_tool_result_xml",type:"covert_tool_invocation",pattern:/<(?:tool_result|function_result)[\s>]/i},{name:"sd_openai_tool_call",type:"covert_tool_invocation",pattern:/"type"\s*:\s*"tool_use"|"tool_calls"\s*:\s*\[/i},{name:"sd_bracket_tool_call",type:"covert_tool_invocation",pattern:/\[(?:TOOL|FUNCTION|CALL)\s*:/i},{name:"sd_double_brace_call",type:"covert_tool_invocation",pattern:/\{\{\s*(?:call|tool|function|invoke)\s*:/i},{name:"sd_invoke_name_attr",type:"covert_tool_invocation",pattern:/<invoke\s+name\s*=/i}],this.COMMAND_INJECTION_PATTERNS=[{name:"shell_injection",pattern:/[;&|`$]|\$\(|\)\s*[;&|]|`[^`]+`/g,severity:50},{name:"command_substitution",pattern:/\$\{[^}]+\}|\$\([^)]+\)/g,severity:50},{name:"pipe_injection",pattern:/\|\s*(cat|rm|curl|wget|nc|bash|sh|exec)/i,severity:55},{name:"path_traversal",pattern:/\.\.[\/\\]|\.\.%2[fF]/g,severity:45},{name:"absolute_path",pattern:/^\/(?:etc|usr|var|tmp|bin|root)/i,severity:40},{name:"oauth_injection",pattern:/authorization_endpoint.*[;&|`$]/i,severity:55},{name:"redirect_manipulation",pattern:/redirect_uri.*[^\w\-_.~:/?#[\]@!$&'()*+,;=%]/i,severity:45},{name:"applescript_injection",pattern:/osascript|do\s+shell\s+script|tell\s+application/i,severity:55},{name:"git_injection",pattern:/--upload-pack|--receive-pack|-c\s+core\./i,severity:50},{name:"git_url_injection",pattern:/ext::|file:\/\/|ssh:\/\/.*@/i,severity:45},{name:"argument_injection",pattern:/\s--[a-z]+=.*[;&|`$]/i,severity:45},{name:"env_injection",pattern:/\bLD_PRELOAD\b|\bPATH\s*=/i,severity:50}],this.SHADOWING_INDICATORS=[{legitimate:"file_reader",suspicious:/file[-_]?read(er)?s?|read[-_]?files?/i},{legitimate:"database_query",suspicious:/db[-_]?query|sql[-_]?query|query[-_]?db/i},{legitimate:"email_sender",suspicious:/send[-_]?emails?|email[-_]?send(er)?/i},{legitimate:"api_caller",suspicious:/call[-_]?api|api[-_]?call(er)?/i},{legitimate:"code_executor",suspicious:/exec[-_]?code|run[-_]?code|code[-_]?run/i}],this.MALICIOUS_SERVER_PATTERNS=[/postmark-mcp.*fake/i,/unofficial/i,/\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}/,/pastebin|gist\.github/i,/temp|tmp|test.*mcp/i],this.config={requireServerSignature:e.requireServerSignature??!1,trustedServers:e.trustedServers??[],blockedServers:e.blockedServers??[],allowDynamicRegistration:e.allowDynamicRegistration??!0,toolAllowlist:e.toolAllowlist??[],toolBlocklist:e.toolBlocklist??[],validateOAuthEndpoints:e.validateOAuthEndpoints??!0,allowedOAuthDomains:e.allowedOAuthDomains??[],detectToolShadowing:e.detectToolShadowing??!0,minServerReputation:e.minServerReputation??30,strictMode:e.strictMode??!1,customInjectionPatterns:e.customInjectionPatterns??[]};for(const s of this.config.trustedServers)this.registeredServers.set(s.serverId,{...s,registeredAt:Date.now(),reputationScore:s.reputationScore??90}),this.serverReputation.set(s.serverId,s.reputationScore??90)}validateServerRegistration(e,s){const i=s||`mcp-reg-${Date.now()}`,t=[];let r=!1,o=!1,a=!1,_=!0,l=50;const{server:n,tools:d,oauth:f,signature:m,timestamp:p}=e;this.isServerBlocked(n.serverId,n.name)&&(t.push("server_blocked"),l=0);const h=this.checkMaliciousPatterns(n);if(h.suspicious&&(t.push(...h.violations),l-=30),this.config.requireServerSignature?!m||!n.publicKey?t.push("missing_server_signature"):(o=this.verifyServerSignature(n,m),o?(r=!0,l+=20):(t.push("invalid_server_signature"),l-=40)):r=!0,this.config.detectToolShadowing){const c=this.detectToolShadowing(d,n.serverId);c.detected&&(a=!0,t.push(...c.violations),l-=50)}for(const c of d)this.config.toolAllowlist.length>0&&!this.config.toolAllowlist.includes(c.name)&&(t.push(`tool_not_in_allowlist: ${c.name}`),_=!1),this.config.toolBlocklist.includes(c.name)&&(t.push(`tool_blocked: ${c.name}`),_=!1),this.detectInjection(c.description).detected&&(t.push(`injection_in_tool_description: ${c.name}`),l-=20);if(f&&this.config.validateOAuthEndpoints){const c=this.validateOAuthConfig(f);c.valid||(t.push(...c.violations),l-=30)}const v=Date.now()-p;v<0?t.push("future_timestamp"):v>300*1e3&&t.push("stale_registration"),!this.config.allowDynamicRegistration&&!this.isTrustedServer(n.serverId)&&t.push("dynamic_registration_disabled"),l=Math.max(0,Math.min(100,l));const g=l<this.config.minServerReputation||this.config.strictMode&&t.length>0||a;return g||this.registerServer(n,d,l),{allowed:!g,reason:g?`Server registration blocked: ${t.slice(0,3).join(", ")}`:"Server registration validated",violations:t,request_id:i,server_analysis:{server_verified:r,signature_valid:o,reputation_score:l,is_shadowing:a,tools_allowed:_},recommendations:this.generateRecommendations(t,"registration")}}validateToolCall(e,s){const i=s||`mcp-call-${Date.now()}`,t=[];let r=!1,o=!0,a=!0,_=!1,l="low";const{toolName:n,serverId:d,parameters:f}=e,m=this.registeredTools.get(n);if(m){r=!0,l=m.riskLevel||"low";const g=this.toolToServer.get(n);g&&g!==d&&(t.push("server_tool_mismatch"),_=!0)}else t.push("tool_not_registered");const p=this.serverReputation.get(d)??0;p<this.config.minServerReputation&&t.push("low_server_reputation"),this.config.toolAllowlist.length>0&&!this.config.toolAllowlist.includes(n)&&(t.push("tool_not_in_allowlist"),o=!1),this.config.toolBlocklist.includes(n)&&(t.push("tool_blocked"),o=!1);const h=this.scanParameters(f);if(h.injectionDetected&&(_=!0,a=!1,t.push(...h.violations)),this.isHighRiskOperation(n,f)&&(l="high",p<70&&t.push("high_risk_low_reputation")),t.length>0){const g=this.serverViolations.get(d)||0;this.serverViolations.set(d,g+t.length);const c=this.serverReputation.get(d)||50;this.serverReputation.set(d,Math.max(0,c-t.length*5))}const v=!r||!o||_||this.config.strictMode&&t.length>0;return{allowed:!v,reason:v?`Tool call blocked: ${t.slice(0,3).join(", ")}`:"Tool call validated",violations:t,request_id:i,tool_analysis:{tool_registered:r,tool_allowed:o,parameters_safe:a,injection_detected:_,risk_level:l},server_analysis:{server_verified:this.registeredServers.has(d),signature_valid:!0,reputation_score:p,is_shadowing:!1,tools_allowed:o},recommendations:this.generateRecommendations(t,"tool_call")}}validateSamplingResponse(e,s){const i=s||`mcp-sampling-${Date.now()}`,t=[];let r=!1,o=!1,a=!1;const _=[],{content:l,serverId:n}=e;(this.serverReputation.get(n)??50)<this.config.minServerReputation&&t.push("low_server_reputation");for(const{name:p,type:h,pattern:v}of this.SAMPLING_ATTACK_PATTERNS)v.lastIndex=0,v.test(l)&&(_.push(p),t.push(`sampling_${h}_${p}`),h==="resource_drain"?r=!0:h==="conversation_hijack"?o=!0:h==="covert_tool_invocation"&&(a=!0));const f=this.detectInjection(l.slice(0,2e3));if(f.detected&&t.push(...f.patterns.map(p=>`sampling_cmd_${p}`)),t.length>0&&n){const p=this.serverReputation.get(n)??50;this.serverReputation.set(n,Math.max(0,p-t.length*10));const h=this.serverViolations.get(n)||0;this.serverViolations.set(n,h+t.length)}const m=r||o||a||this.config.strictMode&&t.length>0;return{allowed:!m,reason:m?`Sampling response blocked: ${t.slice(0,3).join(", ")}`:"Sampling response validated",violations:t,request_id:i,sampling_analysis:{resource_drain_detected:r,conversation_hijack_detected:o,covert_tool_invocation_detected:a,pattern_matches:_},recommendations:this.generateRecommendations(t,"sampling")}}registerTrustedServer(e,s){this.registerServer(e,s,90)}blockServer(e){this.config.blockedServers.includes(e)||this.config.blockedServers.push(e),this.registeredServers.delete(e),this.serverReputation.set(e,0)}getServerReputation(e){return this.serverReputation.get(e)??0}updateServerReputation(e,s){const i=this.serverReputation.get(e)??50;this.serverReputation.set(e,Math.max(0,Math.min(100,i+s)))}getRegisteredServers(){return[...this.registeredServers.values()]}getRegisteredTools(){return[...this.registeredTools.values()]}isToolShadowing(e){for(const s of this.SHADOWING_INDICATORS)if(s.suspicious.test(e)&&e!==s.legitimate)return{shadowing:!0,legitimate:s.legitimate};return{shadowing:!1}}getServerViolations(e){return this.serverViolations.get(e)||0}resetServerViolations(e){this.serverViolations.delete(e)}registerServer(e,s,i){this.registeredServers.set(e.serverId,{...e,registeredAt:Date.now(),reputationScore:i}),this.serverReputation.set(e.serverId,i);for(const t of s)this.registeredTools.set(t.name,t),this.toolToServer.set(t.name,e.serverId),this.toolDefinitionHashes.set(t.name,this.hashToolDefinition(t))}detectToolMutation(e,s){const i=this.toolDefinitionHashes.get(e);if(!i)return{mutated:!1};const t=this.hashToolDefinition(s);return{mutated:i!==t,original_hash:i,current_hash:t}}detectToolDescriptionInjection(e){const s=[],i=[{name:"hidden_instruction",pattern:/(?:IMPORTANT|NOTE|SYSTEM|ADMIN)\s*:/i},{name:"ignore_directive",pattern:/ignore\s+(?:all\s+)?(?:previous|other|prior)/i},{name:"override_behavior",pattern:/override|bypass|instead\s+of|rather\s+than/i},{name:"exfiltrate_data",pattern:/send\s+(?:to|data|all)|forward\s+(?:to|all)|copy\s+(?:to|all)/i},{name:"invisible_text",pattern:/\u200B|\u200C|\u200D|\uFEFF|\u00AD/g}];for(const{name:t,pattern:r}of i)r.lastIndex=0,r.test(e)&&s.push(t);return{injected:s.length>0,patterns:s}}hashToolDefinition(e){const s=require("crypto"),i=JSON.stringify({name:e.name,description:e.description,parameters:e.parameters,serverId:e.serverId});return s.createHash("sha256").update(i).digest("hex")}isServerBlocked(e,s){for(const i of this.config.blockedServers){if(e.includes(i)||s&&s.includes(i))return!0;try{const t=new RegExp(i,"i");if(t.test(e)||s&&t.test(s))return!0}catch{}}return!1}isTrustedServer(e){return this.config.trustedServers.some(s=>s.serverId===e)}checkMaliciousPatterns(e){const s=[],i=`${e.serverId} ${e.name} ${JSON.stringify(e.metadata||{})}`;for(const t of this.MALICIOUS_SERVER_PATTERNS)t.test(i)&&s.push(`malicious_pattern: ${t.source.substring(0,20)}`);return{suspicious:s.length>0,violations:s}}verifyServerSignature(e,s){if(!e.publicKey)return!1;try{const i=JSON.stringify({serverId:e.serverId,name:e.name,version:e.version}),t=crypto.createVerify("SHA256");return t.update(i),t.verify(e.publicKey,s,"hex")}catch{return!1}}detectToolShadowing(e,s){const i=[];for(const t of e){const r=this.toolToServer.get(t.name);r&&r!==s&&i.push(`tool_shadowing: ${t.name} (already registered by ${r})`);const o=this.isToolShadowing(t.name);o.shadowing&&i.push(`suspicious_tool_name: ${t.name} (similar to ${o.legitimate})`)}return{detected:i.length>0,violations:i}}validateOAuthConfig(e){const s=[];if(e.authorizationEndpoint&&(this.detectInjection(e.authorizationEndpoint).detected&&s.push("oauth_authorization_endpoint_injection"),this.config.allowedOAuthDomains.length>0))try{const t=new URL(e.authorizationEndpoint);this.config.allowedOAuthDomains.some(o=>t.hostname.endsWith(o))||s.push(`oauth_domain_not_allowed: ${t.hostname}`)}catch{s.push("invalid_oauth_authorization_url")}return e.tokenEndpoint&&this.detectInjection(e.tokenEndpoint).detected&&s.push("oauth_token_endpoint_injection"),{valid:s.length===0,violations:s}}detectInjection(e){const s=[],i=[...this.COMMAND_INJECTION_PATTERNS,...this.config.customInjectionPatterns.map((t,r)=>({name:`custom_${r}`,pattern:t,severity:50}))];for(const{name:t,pattern:r}of i)r.test(e)&&s.push(t);return{detected:s.length>0,patterns:s}}scanParameters(e){const s=[],i=JSON.stringify(e),t=this.detectInjection(i);t.detected&&s.push(...t.patterns.map(o=>`param_injection_${o}`));for(const[o,a]of Object.entries(e))typeof a=="string"&&a.length>1e4&&s.push(`oversized_parameter: ${o}`);const r=["__proto__","constructor","prototype","eval","exec"];for(const o of Object.keys(e))r.includes(o.toLowerCase())&&s.push(`suspicious_parameter_key: ${o}`);for(const[o,a]of Object.entries(e))typeof a=="string"&&(/^https?:\/\/(?:127\.|10\.|172\.(?:1[6-9]|2\d|3[01])\.|192\.168\.|0\.|localhost|169\.254\.|0\.0\.0\.0|\[?::1\]?)/i.test(a)&&s.push(`ssrf_internal_ip: ${o}`),/^(?:file|gopher|dict|ftp|ldap|ssh|telnet):\/\//i.test(a)&&s.push(`ssrf_dangerous_protocol: ${o}`),/%252e%252e|%c0%ae%c0%ae|%2e%2e%5c|\.\.%255c|\.\.%c0%af|\.\.%c1%9c/i.test(a)&&s.push(`encoded_path_traversal: ${o}`),/\/etc\/(?:passwd|shadow|hosts)|\/proc\/self|\/dev\/(?:null|random)|\.ssh\/|\.env/i.test(a)&&s.push(`sensitive_file_access: ${o}`));return{injectionDetected:s.length>0,violations:s}}isHighRiskOperation(e,s){const i=["execute_code","run_command","shell_exec","eval","file_write","file_delete","database_write","database_delete","send_email","make_payment","transfer_funds","modify_permissions","create_user","delete_user"],t=e.toLowerCase();if(i.some(o=>t.includes(o)))return!0;const r=JSON.stringify(s).toLowerCase();return!!(r.includes("delete")||r.includes("drop")||r.includes("truncate")||r.includes("exec"))}generateRecommendations(e,s){const i=[];return s==="registration"?(e.some(t=>t.includes("signature"))&&i.push("Enable server signature verification for production"),e.some(t=>t.includes("shadowing"))&&i.push("Review tool names for potential shadowing attacks"),e.some(t=>t.includes("oauth"))&&i.push("Configure OAuth domain allowlist"),e.some(t=>t.includes("malicious"))&&i.push("Block suspicious servers and review server sources")):s==="tool_call"?(e.some(t=>t.includes("injection"))&&i.push("Sanitize tool parameters before execution"),e.some(t=>t.includes("reputation"))&&i.push("Only use tools from high-reputation servers"),e.some(t=>t.includes("not_registered"))&&i.push("Register tools before allowing execution")):(e.some(t=>t.includes("resource_drain"))&&i.push("Block this MCP server \u2014 sampling response contains resource exhaustion directives (Unit42/Blueinfy Feb 2026)"),e.some(t=>t.includes("conversation_hijack"))&&i.push("Block this MCP server \u2014 sampling response attempts conversation hijacking via role injection"),e.some(t=>t.includes("covert_tool_invocation"))&&i.push("Block this MCP server \u2014 sampling response embeds covert tool-call syntax")),i.length===0&&i.push(s==="registration"?"Server registration validated successfully":s==="tool_call"?"Tool call validated successfully":"Sampling response validated successfully"),i}}exports.MCPSecurityGuard=MCPSecurityGuard;

package/dist/index.d.ts

@@ -37,7 +37,7 @@ export { CodeExecutionGuard, CodeExecutionGuardConfig, CodeAnalysisResult, Sandb
 export { AgentCommunicationGuard, AgentCommunicationGuardConfig, AgentIdentity, AgentMessage, MessageValidationResult } from "./guards/agent-communication-guard";
 export { CircuitBreaker, CircuitBreakerConfig, CircuitState, CircuitStats, CircuitBreakerResult } from "./guards/circuit-breaker";
 export { DriftDetector, DriftDetectorConfig, BehaviorSample, BaselineProfile, DriftAnalysis, DriftDetectorResult } from "./guards/drift-detector";
-export { MCPSecurityGuard, MCPSecurityGuardConfig, MCPServerIdentity, MCPToolDefinition, MCPServerRegistration, MCPToolCall, MCPSecurityResult } from "./guards/mcp-security-guard";
+export { MCPSecurityGuard, MCPSecurityGuardConfig, MCPServerIdentity, MCPToolDefinition, MCPServerRegistration, MCPToolCall, MCPSamplingResponse, MCPSecurityResult } from "./guards/mcp-security-guard";
 export { PromptLeakageGuard, PromptLeakageGuardConfig, PromptLeakageResult, OutputLeakageResult } from "./guards/prompt-leakage-guard";
 export { TrustExploitationGuard, TrustExploitationGuardConfig, AgentAction, TrustContext, TrustExploitationResult } from "./guards/trust-exploitation-guard";
 export { AutonomyEscalationGuard, AutonomyEscalationGuardConfig, AutonomyRequest, AgentCapabilities, AutonomyEscalationResult } from "./guards/autonomy-escalation-guard";