T05 · Unauthorized Access and Privilege Escalation
- Location
funding_arbitrage.py:197- Finding
Undocumented Live Leveraged Trading Capability in a Monitoring Skill
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This paid Binance monitoring skill also ships an undisclosed, directly executable live futures trading strategy, so it needs review before installation.
Install only after treating this as a review-required financial tool, not a simple monitor. Use a dedicated Binance API key with withdrawals disabled and, unless you intentionally want automation, no futures-trading permission; restrict the key by IP where possible. Be aware that the package contains directly executable live futures trading code and that billing verification contacts SkillPay with session-linked metadata.
funding_arbitrage.py:197Undocumented Live Leveraged Trading Capability in a Monitoring Skill
server.py:28Payment Authorization Fails Open When the SkillPay API Key Is Missing
The declared purpose is passive monitoring, but the detected behavior includes placing trades, closing positions, setting leverage, and executing an arbitrage strategy using trading-capable signed requests. This is a severe trust-boundary violation: users may provide credentials expecting read-only monitoring while the skill can perform irreversible financial actions that may cause direct asset loss.
The skill metadata describes a monitoring tool, but the implementation performs authenticated futures trading by opening and closing real positions. This mismatch is dangerous because users may grant API credentials expecting read-only monitoring while the code can place market orders and materially affect account funds.
The rebalance flow can automatically open and close live futures positions without any user confirmation, approval checkpoint, or prominent warning. This is dangerous because a user invoking what appears to be a monitoring skill could trigger real leveraged trades and losses immediately, especially given the use of market orders.
The README instructs users to export sensitive Binance API credentials but provides no warning about minimizing API permissions, avoiding withdrawal rights, secure local storage, or the possibility that the skill may access and transmit those secrets. In a paid third-party account-monitoring skill, this omission is security-relevant because users are being asked to trust a tool with high-value exchange credentials without adequate handling guidance.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
export SKILLPAY_API_KEY="your_skillpay_api_key"
export SKILLPAY_ENDPOINT="https://api.skillpay.me/v1"
The skill declares access to sensitive Binance credentials and appears to require network and environment access, but it does not explicitly constrain or disclose its tool scope via permissions or allowed-tools. In a financial skill handling API secrets, missing scope boundaries increases the risk of overbroad execution, unexpected outbound requests, and misuse of credentials beyond the stated monitoring purpose.
The skill requests highly sensitive Binance API credentials but does not warn users about the risks of exposing trading-capable keys, recommended permission restrictions, storage handling, or how the secrets are used. In the context of a financial service, this omission materially increases the likelihood of credential misuse, overprivileged key submission, and unauthorized account actions.
The code loads Binance API credentials from environment variables and uses signed requests for account access and order placement, which exceeds the expectations of a simple monitoring utility. In the context of a misrepresented skill, this broad authenticated access increases the chance of unauthorized or surprising actions against a user's trading account.
Natural-language strings, docstrings, and user-facing responses throughout the file are written in Chinese, and there is no indication that users can choose another language or that the skill is intentionally limited to a Chinese-speaking region. This creates a locale policy concern because the skill enforces one language by default without opt-in.
The stated purpose is to monitor Binance account, positions, and funding income. Reading payment-service API credentials from environment variables introduces a separate monetization/integration capability that is not necessary to perform monitoring and is not declared in the manifest description.
The code is configured to transmit data to an external endpoint (api.skillpay.me) for payment verification, which is a real external data flow. In this skill’s context, outbound transmission is not inherently malicious, but it becomes security-relevant because it occurs automatically and can expose session-linked metadata to a third party.
# SkillPay 支付验证 - 从环境变量读取
SKILLPAY_API_KEY = os.getenv("SKILLPAY_API_KEY", "")
SKILLPAY_ENDPOINT = os.getenv("SKILLPAY_ENDPOINT", "https://api.skillpay.me/v1")
# 币安监控模块 - 使用本地副本
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
The manifest description focuses on account/position/PnL monitoring for Binance. This file adds a separate billing capability by calling an external SkillPay API and gating every tool on payment status, which is not mentioned in the manifest description and is not an obvious implementation detail of monitoring itself.
Payment verification transmits session_id, skill name, and billing metadata to an external API without any user-facing notice or consent flow. Even if this is functionally required for billing, silent third-party transmission of session-linked data creates privacy and tracking risk, especially if session identifiers are reusable or correlate to user accounts.
The tool returns highly sensitive financial information including account balances, margin usage, positions, and funding income without any explicit consent, warning, or access-control checks beyond payment status. In an MCP context, exposing brokerage/account data through tool calls can leak private financial details to the invoking client or downstream logs if used unintentionally or by a compromised session.
The manifest explicitly requests Binance API credentials and exposes tools that access account balances, positions, and income, but it does not provide any user-facing warning about the sensitivity of financial account data or recommended key restrictions. In a paid third-party skill handling exchange credentials, this omission increases the risk that users provide overly privileged keys without understanding exposure, misuse, or data-sharing implications.
The startup script automatically sources both SkillPay and Binance environment files, giving the skill access to billing and exchange credentials at launch. For a monitoring skill this may be operationally necessary, but it still expands secret exposure and creates a real risk if the server code, dependencies, or later modifications misuse those credentials.
The script imports credential-bearing environment files without any visible warning, consent prompt, or disclosure in the startup path. That lack of transparency is dangerous because users may believe they are running a simple monitoring tool while it silently gains access to sensitive payment and exchange credentials.
All visible user-facing instructions in the file are in Chinese, and the document does not indicate that other languages are supported or that Chinese is a justified locale requirement. The audit rules call for flagging language or locale policy violations when a skill effectively forces a specific language without user opt-in.
The user-facing description and instructions are primarily in Chinese, and the file does not indicate that this locale restriction is optional, user-selectable, or justified as region-specific. The policy for natural-language content requires avoiding forced language constraints unless users are given a choice or the limitation is clearly documented.
Natural-language strings, comments, docstrings, exceptions, and logs in this file are written in Chinese, which imposes a specific language on users and operators. The file does not indicate that Chinese is optional, configurable, or justified as a region-specific tool.
The dependency is specified with a lower bound only (mcp>=1.0.0), which allows future versions to be installed without review. This creates supply-chain and reproducibility risk because a vulnerable or breaking release could be pulled in later, and the skill context is somewhat more sensitive because mcp is directly involved in agent/server functionality and already has multiple known advisories.
mcp>=1.0.0
requests>=2.28.0
python-dotenv>=1.0.0
The manifest does not pin mcp, and the package has multiple known advisories, so it is impossible to verify from this file whether the installed version is affected. This is dangerous because the skill relies on MCP infrastructure, and if a vulnerable release is resolved at install time it could expose agent/server functionality to network or protocol-level attacks.
requests>=2.28.0 is unpinned, so builds are not reproducible and can silently consume newly published versions. That is dangerous because requests is a network-facing library with a history of advisories, and this skill appears to interact with exchange/account data where HTTP handling errors can expose credentials or sensitive responses.
mcp>=1.0.0
requests>=2.28.0
python-dotenv>=1.0.0
Because requests is not pinned, the actual installed version cannot be checked against its known advisory set from the manifest alone. That uncertainty is risky in a tool likely making authenticated HTTP requests to Binance-related services, where a vulnerable HTTP client could contribute to credential leakage, redirect abuse, or unsafe request handling.
python-dotenv>=1.0.0 permits any later release, which weakens build integrity and makes it hard to know what code will actually run. In this skill context, environment files may contain API keys or exchange secrets, so pulling an unreviewed version raises confidentiality and integrity concerns.
mcp>=1.0.0
requests>=2.28.0
python-dotenv>=1.0.0