Back to skill

Security audit

Siluzan TSO

Security checks for vulnerabilities and agentic risk

Overview

This advertising skill needs review because its optional installers make broad lasting changes and some workflows lack safeguards for sensitive personal data.

Prefer manual installation into one chosen assistant directory, preserving your npm registry and installing Node separately. Explicitly approve target accounts, budgets, live activation, and deletions. Use protected onboarding for identity/banking details and authorize customer-data sharing before uploads. The downloaded CLI and remote dependencies were not included for inspection.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (761)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Installing the CLI and configuring authentication can reasonably support the advertised functions and are not, alone, evidence of mismatch. However, the supplied script also changes shared package-manager configuration, can perform privileged system setup, and invokes siluzan-tso init --global --force for broad assistant registration. These effects extend beyond ordinary execution of the declared advertising and analysis workflows and are not disclosed in the description. The implementation of the invoked CLI is not provided, so the precise files changed by registration and the handling of API credentials cannot be verified. No credential exfiltration or advertising functionality is demonstrated by this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Although the description broadly covers account and financial management, its enumerated capabilities do not disclose withdrawing account funds. The withdraw-submit commands initiate a financially consequential operation distinct from checking balances, recharging wallets, viewing transfer records, or managing account permissions. This is an undeclared capability rather than merely a supporting implementation detail.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 19)May include surrounding context.

md
- 通用纪律与沟通 → `references/core/agent-conventions.md`

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 21)May include surrounding context.

md
> **加载**:按下方路由表 Read **一个**工作流卡片 + **一个**命令/报告 reference 后执行。写操作、报告交付、批量拉数、对用户写结论、或上下文被压缩时,再 Read `references/core/agent-conventions.md`(含 §十四沟通)。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 50)May include surrounding context.

md
要 PDF 就不要 `html_to_pdf`;要了也只转一次,**禁止**再用 pdftotext、PyMuPDF、截图或视觉模型读 PDF/页面。细则见 `agent-conventions.md` §七。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 53)May include surrounding context.

md
时**还没扫完**:原样执行其中的 `resumeCommand`,重复到不再 paused 再交付;禁止交付半截结果,禁止加 `--fresh`。细则见 `references/core/agent-conventions.md` §八。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 57)May include surrounding context.

md
`webUrl`,路径**只**用已 Read 文档中的相对路径表拼接;CLI 输出的授权 URL **原样粘贴**;文档未给路径则**不拼链接**。细则见 `agent-conventions.md`。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 74)May include surrounding context.

md
> 写操作 / 报告交付 / 批量拉数 / 对用户回复 → Read `agent-conventions.md`。找不到行 → 澄清或 Read `intent-routing.md`(勿整本加载 `references/README.md`)。

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 86)May include surrounding context.

md
| 开户(五大媒体)/ 开户进度 | W2 | `references/accounts/open-account-by-media.md`(**首次须列全必填项**;Google 加 `open-account-google-ui.md`) |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 88)May include surrounding context.

md
| 充值 / 钱包 / 转账 / 发票 / 抬头 | W8 | `references/accounts/finance.md` |

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 106)May include surrounding context.

md
| 广告系列/组/广告/关键词 **查询** / 拒审字段 | W3 | `references/google-ads/google-ads-read.md` |

The TikTok workflow submits identity and banking information without an explicit privacy disclosure.

High
Category
Output Handling
Confidence
93% confidence
Finding

The workflow requires collecting a legal representative's identity number, bank account, phone number, and business-license image, then passing sensitive values as CLI arguments without specifying privacy disclosures or secure handling. Account opening provides a legitimate business purpose, and the required pre-submission confirmation mitigates unauthorized submission, but neither addresses informed disclosure of recipients nor potential exposure through chat transcripts, tool logs, or process arguments.

Content

No source excerpt is available for this finding.

The batch-creation workflow creates enabled advertising entities without warning that the changes may go live and incur spend.

High
Category
Output Handling
Confidence
96% confidence
Finding

The recommended workflow explicitly creates an ENABLED ad group and proceeds to add keywords and ads without a local warning or approval checkpoint for live delivery. If the parent campaign is active and the new entities become eligible, this can incur advertising spend; the risk is conditional on account and campaign state, but material in a workflow that directly modifies advertising accounts.

Content

No source excerpt is available for this finding.

The rebuild workflow permits deleting an existing campaign without a deletion-specific warning or confirmation gate.

High
Category
Output Handling
Confidence
96% confidence
Finding

The rebuild instruction explicitly permits deleting an existing campaign before resubmission, without requiring approval of that specific destructive action. Although the broader workflow requires confirmation before creation, approval of a replacement plan does not necessarily authorize removing a live campaign; doing so could interrupt advertising and remove configuration that cannot be restored exactly.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/google-ads/pmax-api.md (reported line 108)May include surrounding context.

md
- WhatsApp **PUT 后 asset id 会变**;后续操作须用响应中的新 `id`
- 创建 Lead Form 后后端自动尝试开启 `SUBMIT_LEAD_FORM` + `GOOGLE_HOSTED` 转化目标

**网关**:`GET /extensionmanagement/pmaxSupportedTypeList`;`POST|PUT|DELETE /extensionmanagement/extension/{accountId}[/{id}]`(Sammamish `ExtensionManagementController.cs`)。

### 图片(推荐流程)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/google-ads/pmax-api.md (reported line 108)May include surrounding context.

md
- WhatsApp **PUT 后 asset id 会变**;后续操作须用响应中的新 `id`
- 创建 Lead Form 后后端自动尝试开启 `SUBMIT_LEAD_FORM` + `GOOGLE_HOSTED` 转化目标

**网关**:`GET /extensionmanagement/pmaxSupportedTypeList`;`POST|PUT|DELETE /extensionmanagement/extension/{accountId}[/{id}]`(Sammamish `ExtensionManagementController.cs`)。

### 图片(推荐流程)

Enhanced Conversions is mandated without a privacy warning about transmitting customer data to Google.

High
Category
Output Handling
Confidence
96% confidence
Finding

The guidance directs all accounts to enable Enhanced Conversions without making customer-data sharing conditional on privacy review, applicable consent, or advertiser authorization. In an operational advertising skill, following this blanket instruction can disclose customer identifiers and CRM-linked conversion information to Google; hashing does not make matchable identifiers anonymous.

Content

No source excerpt is available for this finding.

Enhanced Conversions implementation lacks a warning about sharing customer identifiers and CRM conversion data with Google.

High
Category
Output Handling
Confidence
95% confidence
Finding

The source explicitly describes sending hashed identifiers to Google, so the transfer is not concealed, and its Consent Mode section provides partial mitigation. However, it never gates Enhanced Conversions or CRM uploads on verified permission and applicable privacy requirements, while recommending consent as granted outside the EEA; this can expose linkable customer data, including sensitive lead relationships.

Content

No source excerpt is available for this finding.

The CRM feedback workflow instructs uploading conversion events and hashed emails without a privacy warning.

High
Category
Output Handling
Confidence
98% confidence
Finding

The workflow directs CRM conversion events and hashed email identifiers to Google Ads without requiring an authorization, consent, or applicable privacy-basis check before transfer. Hashing supports identity matching rather than guaranteeing anonymity, so an agent following this operational guidance could disclose linkable customer activity without appropriate approval; the advertising context makes the destination legitimate but does not remove the privacy risk.

Content

No source excerpt is available for this finding.

The correction workflow instructs the agent to upload customer lists without warning about disclosure of customer data to Google.

High
Category
Output Handling
Confidence
96% confidence
Finding

The correction protocol directs the agent to upload fresh customer lists without an explicit authorization or data-handling checkpoint. Customer Match is legitimate in this advertising context, but automatically treating customer-data disclosure to Google as an optimization step could expose personal data beyond the user's intended scope; hashing identifiers does not eliminate that risk.

Content

No source excerpt is available for this finding.

External Script Fetching

High
Category
Supply Chain
Confidence
98% confidence
Finding

The macOS fallback immediately executes an unpinned script from an external hosting endpoint without signature or checksum verification. A compromised endpoint could execute arbitrary commands with the installing user's permissions, potentially exposing local credentials and agent configuration.

Content

Scanner excerpt · scripts/install.sh (reported line 61)May include surrounding context.

sh
brew link --overwrite node@22 2>/dev/null || true
      else
        info "Installing Node.js LTS via install-node.vercel.app..."
        curl -fsSL https://install-node.vercel.app/lts | bash -s -- --yes
      fi
      ;;
    linux)

Chaining Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

The pipeline crosses directly from remotely supplied bytes to root-level shell execution with no verification stage. This is the same underlying issue as the line-67 PE2 findings, not evidence of a separate malicious chaining scheme.

Content

Scanner excerpt · scripts/install.sh (reported line 67)May include surrounding context.

sh
linux)
      if command -v apt-get >/dev/null 2>&1; then
        info "Installing Node.js 22.x via NodeSource (apt)..."
        curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
        sudo apt-get install -y nodejs
      elif command -v yum >/dev/null 2>&1; then
        info "Installing Node.js 22.x via NodeSource (yum)..."

Chaining Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

The pipeline immediately treats a remote response as privileged shell instructions. This duplicates the line-71 remote-execution vulnerability: compromise of the bootstrap source can become full host compromise, even though the apparent purpose is legitimate dependency installation.

Content

Scanner excerpt · scripts/install.sh (reported line 71)May include surrounding context.

sh
sudo apt-get install -y nodejs
      elif command -v yum >/dev/null 2>&1; then
        info "Installing Node.js 22.x via NodeSource (yum)..."
        curl -fsSL https://rpm.nodesource.com/setup_22.x | sudo -E bash -
        sudo yum install -y nodejs
      else
        info "Installing Node.js LTS via install-node.vercel.app..."

External Script Fetching

High
Category
Supply Chain
Confidence
98% confidence
Finding

The Linux fallback pipes a mutable external script directly into bash without verifying the artifact. Compromise of that endpoint could give an attacker arbitrary execution as the installing user, including access to any credentials and files available to that account.

Content

Scanner excerpt · scripts/install.sh (reported line 75)May include surrounding context.

sh
sudo yum install -y nodejs
      else
        info "Installing Node.js LTS via install-node.vercel.app..."
        curl -fsSL https://install-node.vercel.app/lts | bash -s -- --yes
      fi
      ;;
    gitbash)

Dependency Source Redirection

High
Category
Supply Chain
Confidence
100% confidence
Finding

Dependency resolution is redirected away from a canonical default, adds another source, or uses a destination that cannot be resolved statically.

Content

Scanner excerpt · scripts/install.sh (reported line 137)May include surrounding context.

sh
npm config set registry "$NPM_MIRROR"