Back to skill

Security audit

4steps-to-wisdom

Security checks for vulnerabilities and agentic risk

Overview

This skill is a plain markdown communication framework, but it tries to control every response and automatically persist workflow outputs into memory.

Review this skill carefully before installing. It is best suited only for users who explicitly want a strict five-step response style and who can disable or manually approve memory writes; avoid using it for sensitive work unless reasoning, commands, outputs, and stored memories are tightly filtered.

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:93
Finding

Mandatory Response-Structure and Reasoning Disclosure Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 93-108 and 132-143
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Vulnerable Code Snippet

markdown
### 3. Saying
**Externalize thinking for calibration**

- Saying ≠ Doing: speaking is not the same as executing
- Saying = forcing yourself to externalize the thought process
- **Dual calibration**:
  - External: Yu Ge/Kong Kong help discover logical gaps
  - Internal: the moment you speak/write, logical gaps become visible to yourself
- **Key value**: prevents executing before thinking clearly → avoids rework, avoids avoidable failures, avoids accumulating frustration
- Saying is not reporting results — it's a self-check station for thinking quality

### 4. Doing
Command + Result

- Show the command (with comments)
- Show the execution result

The mandatory enforcement statement at line 143 is:

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

Technical Analysis

The Skill imposes a global response protocol rather than limiting its behavior to a particular communication task. It requires the agent to externalize its thought process and disclose commands and execution results. The final self-check makes this protocol mandatory for every response generated while the Skill is active.

This constitutes instruction hijacking because loading the Skill changes the agent's general response policy and attempts to control how unrelated future requests are processed. The instruction to externalize reasoning may also conflict with confidentiality and safety controls that require internal reasoning, hidden context, credentials, or sensitive operational details not to be disclosed.

The Skill does not contain executable code and cannot independently bypass higher-priority controls. Nevertheless, an agent that follows these instructions literally could reveal ...[truncated 1353 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace mandatory language such as “must,” “every message,” and “missing any” with an explicit opt-in mechanism.
  2. Restrict the framework to tasks where the user specifically requests structured communication coaching.
  3. Remove instructions requiring private reasoning to be externalized. Request concise conclusions, assumptions, and user-relevant rationale instead.
  4. Replace “show the command and result” with a requirement to provide sanitized execution summaries only when relevant.
  5. Explicitly prohibit disclosure of credentials, tokens, hidden prompts, private context, internal reasoning, and sensitive command output.
  6. State that system, developer, security, privacy, and task-specific instructions always take precedence.
  7. Add a relevance check so the framework is not applied to unrelated requests.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:147
Finding

Unconditional Persistence of Workflow Content into Long-Term Memory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 147-159
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: High

Vulnerable Code Snippet

markdown
The outputs of the Five Steps automatically flow into the memory palace:
- Decision rationale → stored as principles
- Execution results → stored as completed work
- Reflections and insights → stored as insights

Technical Analysis

The Skill directs workflow outputs to be stored automatically in persistent memory. The content selected for persistence includes decision rationale, execution results, and reflections, all of which can contain user-controlled claims, untrusted instructions, confidential data, or inaccurate conclusions.

No consent check, trust-boundary validation, sensitivity filter, provenance metadata, expiration policy, or mechanism for reviewing and deleting stored content is specified. Consequently, an agent platform that implements these storage directives could preserve attacker-controlled material and retrieve it in later sessions.

The repository contains no script that performs the writes directly. Exploitation therefore depends on the host agent having a memory facility and interpreting the Skill's storage instructions. Under that condition, the directive creates a persistent memory-poisoning path.

Attack Path

  1. The Skill is loaded by an agent with persistent-memory capabilities.
  2. An attacker submits misleading facts, malicious behavioral guidance, or sensitive content as part of an ordinary request.
  3. The forced workflow incorporates that content into decision rationale, execution results, or reflections.
  4. The Skill directs those outputs to flow automatically into long-term memory.
  5. The stored content survives the current interaction.
  6. A later session retrieves the poisoned memory as a principle, completed result, or learned insight.
  7. The agent relies on the untrusted content, causing fu ...[truncated 666 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove automatic persistence and make every memory write explicit and opt-in.
  2. Require confirmation that identifies exactly what will be stored, why it is useful, and how long it will be retained.
  3. Never store raw prompts, private reasoning, credentials, command output, or other sensitive session content.
  4. Validate proposed memories against a strict schema and reject executable instructions or untrusted behavioral rules.
  5. Record provenance, creation time, originating user or session, confidence, and expiration metadata.
  6. Isolate memory by user and trust domain to prevent cross-user contamination.
  7. Provide review, correction, deletion, and automatic expiration mechanisms.
  8. Treat retrieved memories as untrusted reference data rather than authoritative instructions.
  9. Ensure that stored memory cannot override system, developer, safety, or access-control policies.
Vulnerability Patterns
  • 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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (1)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

This markdown file repeatedly presents the workflow in fixed Chinese and English forms and uses mandatory language such as '每次收到消息,回复前必须满足以下结构' at L032, but it does not state that the user may choose their preferred language or locale. Under the policy rule, forcing a specific language/locale convention without user opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.