Back to skill

Security audit

Skill Compounding

Security checks for vulnerabilities and agentic risk

Overview

This skill is not malicious, but it asks agents to persistently create and rewrite skill instructions with broad triggers and no required user approval.

Install only if you want an agent to help create and revise reusable skill documentation. Before use, require explicit approval for every created or modified skill file, avoid letting it write to SOUL.md automatically, review full diffs, and do not preserve guidance derived from untrusted task content without sanitization.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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 (1)

T02 · Agent Memory Poisoning

Warning
Location
SKILL.md:72
Finding

Persistent Agent Guidance Can Be Modified Without Mandatory User Approval

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 72-81
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: Medium

Relevant instruction excerpt — faithful English translation of the reviewed Markdown:

markdown
### Phase C: Register in the Skill Index

Add the new Skill to `skills/quick-reference.md` if it exists, or declare it
in the Skill hierarchy in `SOUL.md`.

### Phase D: Continuous Iteration

Every time this Skill is used:
- If a problem is found, update `SKILL.md`
- If a new scenario is found, extend `SKILL.md`
- If a better solution is found, rewrite the corresponding section

Technical Analysis

The Skill instructs the agent to write generated guidance into persistent files, including skills/quick-reference.md, other SKILL.md files, and particularly SOUL.md. These files may influence behavior in later tasks or sessions.

The workflow does not require explicit user confirmation, trusted-source validation, content sanitization, or review of a proposed diff before persistence. Because the decision to preserve a pattern is based on task content and the agent's assessment, untrusted instructions presented during an ordinary task could be misclassified as reusable guidance and stored for future use.

This is a memory-poisoning risk rather than operating-system persistence: the package does not install services, scheduled tasks, executable hooks, or startup components. No executable code, remote payload retrieval, dependency installation, credential access, or network behavior was found.

Attack Path

  1. An attacker supplies task content containing a seemingly reusable workflow with embedded hostile or policy-altering instructions.
  2. The agent completes the task and evaluates the workflow under the Skill's broad reuse criteria.
  3. The agent decides that the attacker-influenced workflow should be preserved.
  4. Following Phase C, the agent registers the genera ...[truncated 857 chars]
Remediation
View remediation

Remediation Suggestions

  1. Require explicit user approval before creating, registering, extending, or rewriting any persistent Skill content.
  2. Do not permit this workflow to modify SOUL.md; use a dedicated, narrowly scoped Skill registry instead.
  3. Generate and display a complete diff before applying any persistent change.
  4. Record the source and trust level of proposed reusable guidance.
  5. Reject content containing instruction overrides, credential requests, unrelated tool commands, external payload retrieval, or attempts to weaken security controls.
  6. Separate untrusted task data from agent instructions and ensure copied examples cannot become executable directives.
  7. Apply least-privilege filesystem permissions so the Skill can write only to an approved staging directory.
  8. Require human review before promoting staged content into an active Skill index or other automatically loaded instruction file.
  9. Maintain version history and provide a rollback mechanism for all persisted Skill changes.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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
Findings (3)

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 1)May include surrounding context.

md
# skill-compounding

> **借鉴来源**:Multica (multica-ai/multica) 的 Reusable Skills 机制
>

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

该 markdown 文件定义了技能触发条件,但其中“成功解决了一个非平凡问题”“发现了一个通用的解决方案”“某个模式出现了 2 次以上”等表述缺乏可操作的边界,何时算“非平凡”“通用”或“模式出现”并不明确。这会使技能在大量普通工作对话中都可能被认为应当激活,且文档未提供负例或排除条件来收窄范围。

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

“把这个流程记下来,以后复用”等说法虽然与技能主题相关,但仍可能在一般协作或笔记场景中自然出现,从而与非该技能的普通请求发生重叠。该部分只给出正向示例,未说明必须满足的上下文条件或不应触发的类似说法。

Content

No source excerpt is available for this finding.