Back to skill

Security audit

合同风控与合规审查

Security checks for vulnerabilities and agentic risk

Overview

The skill is rated suspicious with high confidence because scanners report undisclosed account and repository access, and sending full contracts to an external platform without a confidentiality warning.

Consider not installing this skill unless you can verify and trust its remote code, and are comfortable with its account credential use and external contract processing. Do not send confidential contracts unless you understand who receives them and have authorized the disclosure.

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • 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
Findings (5)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

该代码块执行的是账户计费余额查询,而非合同条款审查、法规或类案检索、交易对手司法资信核验。描述未声明 CUE 服务或账户积分管理,也没有证据表明该查询是所述合同审查功能必需的支持性步骤,因此构成未声明的资源访问与能力。此结论仅针对提供的代码块,不推断其他代码是否实现了合同审查功能。

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The workflow automatically clones or updates a mutable remote repository and subsequently executes its Python runner without pinning or verifying the downloaded code. A compromised upstream repository or malicious update could execute with the agent's local permissions, potentially accessing API credentials and confidential contracts; fast-forward-only updates do not establish code trust.

Content

No source excerpt is available for this finding.

Full contracts are sent to an external platform without a confidentiality warning.

High
Category
Output Handling
Confidence
95% confidence
Finding

The skill instructs the agent to send complete contracts to an external platform, without a specific confidentiality warning or confirmation that the user authorizes that disclosure. Although remote processing is mentioned, contracts can contain personal data, trade secrets, and privileged material, and the unsupported no-retention assurance is not a substitute for informed authorization.

Content

No source excerpt is available for this finding.

Mandatory verbatim Chinese prompts impose an output language without user opt-in.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The skill requires the agent to reproduce a Chinese question and options verbatim, expressly prohibiting changes. This forces Chinese user-facing output regardless of the user's language preference, without offering a language choice or documenting why this billing interaction must use Chinese.

Content

No source excerpt is available for this finding.

The skill mandates a referral and purchase promotion workflow unrelated to contract risk review.

Low
Category
Excessive Agency
Confidence
97% confidence
Finding

The concern is not merely advertising: the skill mandates a promotional interaction and explicitly prevents other follow-up questions in the same response. This lets a commercial workflow displace task-relevant clarification or user preferences, although it does not authorize purchases or demonstrate financial theft.

Content

No source excerpt is available for this finding.