Back to skill

Security audit

Vibe Coding Learning

Security checks for vulnerabilities and agentic risk

Overview

This learning-note skill is mostly coherent, but it needs Review because it can broadly read project code, prior notes and conversation history, write persistent learning files, and run ambiguous local scripts.

Install only if you are comfortable with the skill reading project files, learning notes, and selected conversation history, then saving structured notes under `learning-notes/`. Before using it, ask the agent to show the files or conversations it plans to read, disable or avoid deep-mode WebSearch for confidential projects, and do not allow the referenced local Python scripts to run unless you know exactly which files they resolve to.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (10)

Context Leakage

High
Category
Data Exfiltration
Confidence
96% confidence
Finding

The 'Collect Context' step explicitly instructs scanning the project directory and reading key source files to understand what was built. In this skill context, that is more dangerous because the feature is framed as educational summarization, yet it can ingest broad local context, including unrelated or sensitive files, without strong scoping or consent.

Content

Scanner excerpt · SKILL.md (reported line 134)May include surrounding context.

md
> 进度输出(仅 standard/deep 模式): "正在分析你的代码结构..."

**Step 1 — Collect Context**
- Scan the project directory to identify files created or modified during this session
- If user provided `coding_context`, use it as the primary description
- Read the key source files to understand what was built

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger phrases include broad, common language such as requests to explain code, review, or summarize learning. This can cause unintended invocation, leading the skill to scan files or write notes when the user may have wanted a narrower, non-persistent response.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill directs broad project scanning and persistent note writing without an upfront warning that local files will be read and modified. This creates a meaningful privacy and integrity risk because users may unknowingly expose unrelated project content or permit writes to their workspace.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill instructs local subprocess execution via python scripts to analyze session complexity and validate structure. Spawning local commands increases the attack surface and may execute repository-controlled code or scripts that are outside the narrowly described note-generation function.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill instructs web searches using session-derived knowledge points but does not warn users that derived topics may be sent to external services. Even summarized topics can reveal private project details, internal technologies, or confidential work contexts.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

Mode 4 directs the skill to scan accumulated notes, domain indexes, and knowledge cards to build interview output from historical material. While related to the educational purpose, it still broadens data access to a larger corpus than the current interaction and can reveal past sensitive notes or project details.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest positions the skill as learning-note generation from coding sessions, but this section expands behavior into prompt/history analysis of user-AI conversations. That is a material scope expansion because it causes the skill to process additional sensitive data types beyond what a user would reasonably expect from the declared purpose.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Mode 5 instructs reading and analyzing recent conversation history, including imported web chat data. This is dangerous because it processes potentially sensitive user content beyond the immediate request scope and could surface secrets, personal data, or unrelated conversations into the skill's analysis output.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The documented behavior goes beyond summarization into knowledge-base maintenance actions such as inbox triage, backlink insertion, weekly synthesis, and health checks. These actions can modify persistent user data and broaden access to unrelated notes, creating a mismatch between declared purpose and actual authority.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The writing rules state "Write in plain Chinese," which imposes a specific language policy on outputs. There is no accompanying opt-in or alternative language option for users who may prefer another language.

Content

No source excerpt is available for this finding.