Back to skill

Security audit

沟通五部曲

Security checks for vulnerabilities and agentic risk

Overview

This skill is a workflow document, but it broadly forces every response into a fixed process and directs automatic persistence of conversation-derived content without clear consent or retention controls.

Review this skill carefully before installing. It should only be used where you want a rigid five-step response style and where memory writes are disabled or explicitly confirmed by the user. Avoid using it with sensitive work unless the host agent prevents automatic storage of conversation content and command output.

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
Findings (2)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:169
Finding

Mandatory Global Response Workflow Hijacks Agent Behavior

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:169
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Critical

Vulnerable Code

text
Before sending, verify all five modules are present. Missing any → complete before sending.

Technical Analysis

The Skill imposes an unconditional response requirement rather than limiting the workflow to situations where a user explicitly requests it. Once loaded, the instruction requires every response to contain all five prescribed modules.

Other portions of the documented workflow require the agent to externalize its thinking and display commands and execution results. Applying those requirements globally can override task-appropriate response behavior and conflict with higher-level confidentiality or least-disclosure expectations.

The package contains only documentation and does not include executable code. Consequently, this issue affects agent instructions and current-session behavior rather than directly granting operating-system privileges.

Attack Path

  1. The agent loads SKILL.md as an active Skill.
  2. The unconditional pre-send rule becomes part of the agent's current instructions.
  3. A user submits a task containing sensitive context or requiring tool execution.
  4. The Skill directs the agent to apply every workflow module, including externalized reasoning and command/result reporting.
  5. Internal operational details, commands, or sensitive command output may consequently be disclosed in the response.

Impact Assessment

The Skill can influence all responses produced while it is active, altering the agent's output structure and task execution behavior. It may cause unnecessary disclosure of reasoning, commands, execution results, or sensitive contextual information.

No direct privilege escalation, arbitrary code execution, or host compromise was identified. The affected scope is the agent's current session and any inform ...[truncated 36 chars]

Remediation
View remediation

Remediation Suggestions

  • Make the five-step workflow explicitly opt-in and activate it only when requested by the user.
  • State that system, developer, security, confidentiality, and tool-use constraints take precedence.
  • Do not require disclosure of private reasoning or hidden intermediate analysis. Request concise conclusions and relevant rationale instead.
  • Replace raw command and result disclosure with sanitized execution summaries.
  • Permit modules to be omitted when they are unnecessary, unsafe, or incompatible with the active task.
  • Add explicit filtering requirements for credentials, tokens, personal data, filesystem paths, and other sensitive output.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:182
Finding

Automatic Persistence of Conversation-Derived Content Enables Memory Poisoning

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:182-185
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: High

Vulnerable Code

text
The outputs of the Five Steps automatically flow into the memory palace:
- Decision rationale → stored in 热/原则 (principles)
- Execution results → stored in 温/done (completed work)
- Reflections and insights → stored in 热/领悟 (insights)

Technical Analysis

The Skill directs the agent to persist decision rationale, execution results, and reflections automatically. It does not require user consent, content validation, provenance tracking, sensitivity filtering, expiration, or isolation between users and sessions.

Conversation-derived material can include attacker-controlled assertions, embedded instructions, secrets, misleading conclusions, and unsafe execution results. If the referenced memory system is available, automatic storage can cause that untrusted content to be retrieved in later sessions and treated as trusted historical knowledge.

The repository does not contain the memory implementation, so an actual persistent write depends on the host agent providing the referenced integration. The vulnerability is nevertheless present in the Skill instruction because it mandates the unsafe persistence behavior.

Attack Path

  1. The agent loads the Skill in an environment connected to persistent memory.
  2. An attacker submits crafted content designed to appear as a principle, decision rationale, completed result, or learned insight.
  3. The five-step process incorporates the attacker-controlled material into its output.
  4. The automatic persistence instruction stores that material in the memory palace.
  5. A later session retrieves the stored content through vector or Wiki search.
  6. The agent relies on the poisoned content when making subsequent decisions, allowing the attacker's influence to persist across sessions.

Impact Assessment

If pe ...[truncated 642 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove automatic memory writes and require explicit, informed user approval for each persistence operation.
  • Treat all conversation-derived content as untrusted before storage.
  • Store only narrowly scoped, structured facts rather than free-form instructions, reasoning, or raw execution results.
  • Reject secrets, credentials, personal data, executable instructions, and unverified claims.
  • Attach provenance, author identity, timestamps, confidence levels, and source-session identifiers to every record.
  • Isolate memory by user, workspace, and task to prevent cross-session or cross-tenant contamination.
  • Apply retention periods and provide review, correction, and deletion controls.
  • Ensure retrieved memory is treated as untrusted context rather than authoritative instructions.
  • Require confirmation before retrieved content can alter security-sensitive decisions or trigger tool execution.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The skill mandates a rigid five-part reply structure ('每次收到消息,回复前必须满足以下结构') without user opt-in, which can override user preferences and system expectations. While not directly enabling code execution or data exfiltration, it can degrade safe operation by coercing responses into a fixed format even when inappropriate.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill states that outputs 'automatically flow into the memory palace' and are stored in persistent categories, but it provides no user-facing notice, consent flow, or retention limits. This creates a privacy risk because sensitive user content, decision rationale, and reflections may be retained without informed consent.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instructions explicitly direct automatic retention of user interaction artifacts into persistent memory stores ('决策过程', '执行结果', '复盘收获'). Persistent recording of conversational content increases the chance of privacy leakage, secondary misuse, and long-term exposure of sensitive operational or personal information.

Content

No source excerpt is available for this finding.