Back to skill

Security audit

WTT

Security checks for vulnerabilities and agentic risk

Overview

This is a real WTT integration, but it has serious under-disclosed credential, encryption, and automatic remote-task execution risks.

Review this plugin carefully before installing. Use it only with a trusted WTT backend and narrowly scoped agent token, avoid custom or HTTP cloud URLs, assume E2E protection is not a strong confidentiality boundary, and rotate tokens if it has already handled untrusted media URLs or run with permissive local config permissions.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (7)

other

Error
Location
src/ws-client.ts:410
Finding

Remote WTT server can retrieve the end-to-end encryption master key

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/channel.ts:1291
Finding

Agent bearer token is sent to arbitrary inbound media URLs

Content
View full analysis
{ const headers: Record = { Accept: "image/*,*/*;q=0.8", }; if (params.account.token) { headers.Authorization = `Bearer ${params.account.token}`; headers["X-Agent-Token"] = params.account.token; } const controller = new AbortController(); const timeout = setTimeout(() => controller.abort(), Math.max(1000, params.timeoutMs)); let response: Response; try { response = await fetch(params.url, { method: "GET", headers, redirect: "follow", signal: controller.signal, }); } finally { clearTimeout(timeout); } ``` The function is invoked for inbound message URLs without origin validation: ```ts const downloaded = await downloadInboundMediaToLocal({ url, account: params.account, maxBytes, timeoutMs, }); ``` ### Technical Analysis The media URL is derived from inbound messaging data, but the downloader attaches both the bearer `Authorization` header and `X-Agent-Token` to every destination. It does not enforce an origin allowlist, require HTTPS, reject loopback or private-network addresses, or validate the destination after redirects. The request also uses `redirect: "follow"`, so the effective destination can differ from the initially supplied URL. Even where a runtime strips sensitive headers on some cross-origin redirects, the initial arbitrary destination still receives them, and the request remains an SSRF primitive. ### Attack Path 1. An attacker sends an inbound WTT message containing a media URL hosted on an attacker-controlled server. 2. The plugin extr ...[truncated 871 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/e2e-crypto.ts:69
Finding

Message encryption uses unauthenticated AES-CTR

Content
View full analysis
{ const counter = await hkdfNonceWeb(key, contextId); const aesKey = await crypto.subtle.importKey("raw", key.buffer as ArrayBuffer, { name: "AES-CTR" }, false, [ "encrypt", ]); const ciphertext = await crypto.subtle.encrypt( { name: "AES-CTR", counter: counter.buffer as ArrayBuffer, length: 128 }, aesKey, plaintext.buffer as ArrayBuffer, ); return new Uint8Array(ciphertext); } async function decryptWeb(key: Uint8Array, ciphertext: Uint8Array, contextId: string): Promise { const counter = await hkdfNonceWeb(key, contextId); const aesKey = await crypto.subtle.importKey("raw", key.buffer as ArrayBuffer, { name: "AES-CTR" }, false, [ "decrypt", ]); const plaintext = await crypto.subtle.decrypt( { name: "AES-CTR", counter: counter.buffer as ArrayBuffer, length: 128 }, aesKey, ciphertext.buffer as ArrayBuffer, ); return new Uint8Array(plaintext); } ``` The implementation explicitly documents the absence of authentication: ```ts /** * Decrypt ciphertext bytes back to plaintext string. * AES-CTR has no authentication — caller must handle garbled output gracefully. */ ``` ### Technical Analysis AES-CTR is a stream-like encryption mode and does not provide message integrity or authenticity. Modifying a ciphertext bit causes a predictable modification to the corresponding plaintext bit. Because no MAC or authentication tag is generated or verified, altered ciphertext is accepted and returned as plaintext. The derivation also uses the fixed salt `wtt-e2e-shared`, and the accepted `agentId` parameter is ignored. Consequently, reuse of an E2E password produces the same master key across accounts and contexts. ### ...[truncated 744 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/runtime/task-executor-persistence.ts:256
Finding

Task queue persistence serializes bearer tokens in plaintext

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
bin/openclaw-wtt-bootstrap.mjs:129
Finding

Bootstrap writes plaintext tokens and backup copies without permission hardening

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
src/ws-client.ts:257
Finding

Unrestricted cloud URL permits credential transmission to insecure or attacker-controlled endpoints

Content
View full analysis
{ this.log("info", "WebSocket connected, authenticating..."); this.reconnectAttempts = 0; // Send auth as first message this.send(JSON.stringify({ action: "auth", token: this.account.token })); this.startHeartbeat(); this.onConnect?.(); }; ``` ### Technical Analysis No validation requires `cloudUrl` to use HTTPS or belong to a trusted origin. The textual replacement maps `http://` to `ws://`, causing the token to traverse an unencrypted WebSocket connection. A deliberately attacker-controlled HTTPS endpoint is also accepted and receives the token. Because the same configurable origin is used by authenticated REST calls, unsafe configuration can expose credentials through both WebSocket and HTTP code paths. ### Attack Path 1. A malicious instruction, configuration import, or operator mistake sets `cloudUrl` to `http://attacker.example` or another untrusted endpoint. 2. `resolveWsUrls` constructs `ws://attacker.example/ws/`. 3. The client connects without transport encryption. 4. On connection, it sends `{ action: "aut ...[truncated 363 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
src/runtime/task-executor.ts:380
Finding

Remote task descriptions are inserted into automatically triggered Agent prompts

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (90)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Unlike many other TP4 findings, this one partly maps to real risk in the documentation: the skill instructs users to install/enable a plugin, restart the gateway, and run bootstrap commands, but does not disclose the sensitivity and side effects of these operations. In a security-sensitive agent ecosystem, under-describing config modification, package trust, and service restarts can mislead users into executing impactful commands without adequate review.

Content

No source excerpt is available for this finding.