Back to skill

Security audit

Spider Web Trigger Network

Security checks for vulnerabilities and agentic risk

Overview

This skill is a real skill router, but it needs review because it can scan all installed skills, route into other skills, and expose a local dashboard with weak safety controls.

Review before installing. Use this only in a trusted local skill set, avoid starting the dashboard on a network-facing host, and do not let it automatically load matched skills without user confirmation. Treat the trigger database as sensitive because it reveals installed skills and routing metadata.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/server.py:187
Finding

Stored Cross-Site Scripting Through Unescaped Skill Metadata

Content
View full analysis
${s.total_skills}
已索引技能
${s.total_triggers}
触发词总数
${s.unique_triggers}
唯一触发词
${s.overlap_triggers}
重叠触发词
${s.avg_per_skill}
平均触发词/技能
${s.network_density||'N/A'}%
网络密度
`; } function renderSkills(d) { const skills = d.skills || {}; const sorted = Object.entries(skills).sort((a,b) => b[1].length - a[1].length); document.getElementById('skill-list').innerHTML = sorted.map(([name,triggers]) => `
🕸️ ${name} ${triggers.length} 触发词
`).join(''); } async function testMatch() { const query = document.getElementById('query-input').value.trim(); if (!query) return; const r = await fetch('/api/match', { method: 'POST', headers: {'Content-Type': 'application/json'}, body: JSON.stringify({query, mode: 'auto', fuzzy: fuzzyMode, top_k: 10}) }); const result = await r.json(); const container = document.getElementById('match-results'); if (!result.matches || result.matches.length === 0) { container.innerHTML = `
❌ 未找到匹配技能
${result.suggest ...[truncated 2855 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/server.py:284
Finding

Unauthenticated Management and Data APIs With Permissive Cross-Origin Access

Content
View full analysis
0 else b'{}' try: data = json.loads(body) except: data = {} if path == '/api/match': self._handle_match(data) elif path == '/api/reindex': self._handle_reindex() else: self.send_error(404) def _handle_match(self, data): if not self.engine: self.engine = SpiderMatchEngine() query = data.get('query', '') mode = data.get('mode', 'auto') fuzzy = data.get('fuzzy', False) top_k = data.get('top_k', 10) result = self.engine.match(query, mode=mode, top_k=top_k, fuzzy=fuzzy) self._serve_json(result) def _handle_reindex(self): try: db = index_all_skills() save_db(db) self.engine = SpiderMatchEngine() # reload self._serve_json(db) except Exception as e: self._serve_json({"error": str(e)}, 500) def _get_full_data(self): db = load_db() if not db: return {"error": "No database found"} if self.engine and self.engine.db: stats = self.engine.get_stats() db['meta']['network_density'] = stats.get('network_density' ...[truncated 3453 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
scripts/index_triggers.py:171
Finding

Untrusted Skill Metadata Can Poison Automatic Agent Routing

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (32)

Vague Triggers

Critical
Category
Not specified by scanner
Confidence
99% confidence
Finding

Using file extensions and broad tokens like “技能” as triggers makes accidental invocation highly likely whenever users discuss documents or skills generally. This is risky in a router because passive mention of common artifact types can unintentionally activate analysis workflows.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The code accurately implements only the indexing portion of the declared system: scanning installed skills, parsing YAML frontmatter descriptions, extracting trigger words, building a unified trigger database, reverse index, and reporting results. However, the description claims a broader routing and matching system with serial/parallel matching, semantic matching, fuzzy matching, AND/OR logic, and a web admin panel. None of those capabilities appear in this code chunk. The filesystem access to ~/.workbuddy/skills/ is consistent with the indexing purpose, and there is no obvious unrelated or malicious behavior. Still, because the declared description substantially overstates the implemented capabilities of this code chunk, this is a description-behavior mismatch.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger phrases are very broad and overlap with ordinary conversation such as asking which skill to use or requesting help with matching. This can cause unintended invocation of a high-privilege meta-skill that scans other skills and may launch follow-on skills, increasing the chance of accidental capability escalation or misrouting.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 69)May include surrounding context.

md
这会在 `scripts/trigger_db.json` 生成完整的触发词数据库。

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 158)May include surrounding context.

md
这会在 `scripts/trigger_db.json` 生成完整的触发词数据库。

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 182)May include surrounding context.

md
每个技能的触发词定义在 `SKILL.md` 的 `description` 字段中,格式:

Lp1

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'network' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.

Content

No source excerpt is available for this finding.

Unvalidated Output Injection

High
Category
Output Handling
Confidence
96% confidence
Finding

The dashboard renders server-supplied values such as m.skill, matched_triggers, and suggestion directly into innerHTML and also interpolates m.skill into an onclick attribute without escaping. If any indexed skill metadata or match result contains attacker-controlled HTML/JavaScript, opening the dashboard can trigger stored or reflected XSS in the browser, which is especially relevant because this system scans installed skills and displays their content-derived data.

Content

Scanner excerpt · scripts/server.py (reported line 226)May include surrounding context.

python
return;
  }
  
  container.innerHTML = result.matches.map((m,i) => {
    const cls = m.score >= 10 ? 'high' : m.score >= 5 ? 'medium' : 'low';
    return `<div class="match-result ${i===0?'best':''}" onclick="document.getElementById('query-input').value='${m.skill}';testMatch()">
      <span class="skill-name">#${i+1} ${m.skill}</span>

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Triggers like common words such as “包含” and “技术” are so broad that ordinary conversation may activate the wrong skill. In a router, this is dangerous because false activations can redirect user input to unintended downstream skills, including ones with higher privileges or side effects.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The trigger set includes highly generic terms with little contextual anchoring, making accidental matches likely in unrelated user requests. Because this file powers skill routing, ambiguous triggers directly increase the chance of invoking the wrong capability.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Bookkeeping triggers like everyday verbs and phrases such as spending or receiving money overlap heavily with normal chat. This can cause silent or repeated activation of a financial-recording skill, leading to incorrect handling of personal finance content and unintended downstream actions.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

Requests like “帮我分析数据” and similar broad analysis phrases are common across many workflows and are not unique to one skill. A routing engine using these triggers can easily send users to the wrong analyzer, especially when other skills also process reports, products, or metrics.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Broad shopping phrases such as asking whether a product is worth it or to analyze a link are likely to appear in ordinary conversation. In this router, that creates a high risk of unintended ecommerce-skill invocation and improper handling of arbitrary links or product discussions.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

Generic analysis phrases such as product analysis and conversion analysis are not unique to this interest-analysis skill and overlap with several other commerce skills. In a central router, these collisions can systematically misroute user requests and surface the wrong business logic.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Terms such as “每日” are extremely generic and have almost no direct connection to nutrition intent by themselves. These triggers can match a large volume of normal conversation, causing widespread false routing into the nutrition skill.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Triggers like “全国” and “分数匹配” are far too generic to safely indicate gaokao-volunteer counseling. In a router, these could hijack unrelated educational, analytics, or national-scope queries and misdirect users to the wrong skill.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The anonymous or opaque skill entries use triggers like “create,” “remove,” “use,” and similar generic command words with no clear scope. In a skill-routing system, this is especially dangerous because users can be redirected into unknown skills through ordinary commands, potentially reaching hidden capabilities with destructive side effects.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Placeholder-style patterns like “XX教程” and broad help phrases are open-ended and can absorb many unrelated requests. In a routing engine, such templates effectively act as catch-alls and can crowd out more appropriate skills.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Phrases such as "怎么用XX", "XX教程", and "帮我学XX" are template-like and extremely broad, overlapping with ordinary requests for help about any topic. The manifest gives no negative examples or constraints for when this skill should or should not activate.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The README states in Chinese that the system routes WorkBuddy skills based on trigger words, and the matching design prominently targets Chinese input and Chinese trigger labeling. There is no indication that users may choose another language or that the locale restriction is explicitly justified as region-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README advertises a web management panel with endpoints that expose the complete indexed database and permit reindexing, but it does not document any authentication, access controls, or privacy warnings. In this skill’s context, the indexed data is derived from all installed skills and their trigger metadata, so an exposed panel or API could leak the local skill inventory and enable unauthorized administrative actions such as reindexing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The display name, description, trigger phrases, and examples are predominantly Chinese, and the matching design explicitly emphasizes Chinese semantic matching, while the document does not offer a language choice or state that the skill is intentionally region-specific. This can conflict with language/locale policy expectations when users have not opted into Chinese-only behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill describes convenient routing features but does not clearly warn that it scans all installed skills, writes a persistent JSON index, and can start a local web management panel. Missing disclosure reduces informed consent and can expose metadata about installed capabilities or make local services available without users appreciating the privacy and attack-surface implications.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
80% confidence
Finding

This skill has broad capabilities including Write, Edit, Bash, Glob, Grep, Skill, and task-management tools, while its purpose is to scan the environment, build a database, and route to other skills. In a meta-router context, persistent file writes and shell execution increase the blast radius of accidental or adversarial triggering, because the skill can enumerate local content, store session-derived data, and chain into additional skills.

Content

Scanner excerpt · SKILL.md (reported line 16)May include surrounding context.

md
agent_created: true
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Glob

Dynamic import via __import__()

Medium
Category
Dangerous Code Execution
Confidence
75% confidence
Finding

Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.

Content

Scanner excerpt · scripts/match_engine.py (reported line 46)May include surrounding context.

python
self.idf = {}
        for trigger, skills in self.reverse_index.items():
            df = len(skills)
            self.idf[trigger] = __import__('math').log(total / max(df, 1)) + 1.0
        # Cache for fast lookup
        self._max_idf = max(self.idf.values()) if self.idf else 1.0