Back to skill

Security audit

deep-skill-finder

Security checks for vulnerabilities and agentic risk

Overview

This skill needs review because its privacy promises conflict with data sharing, and flaws can upload files prematurely or overwrite other skills on Windows.

Defer installation until archive containment and upload validation order are fixed. Keep weekly personalization disabled unless you explicitly approve sharing work-derived keywords. Use generalized search queries and inspect complete feedback and media before submission; redaction does not protect screenshot or video contents. Review downloaded skills separately. Remote package contents, server retention, and deletion behavior were not verified.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

other

Warning
Location
scripts/deep_skill_search.py
Finding

Weekly recommendations transmit context-derived keywords despite local-only privacy assurances

Content
View full analysis

Impact

Keywords can reveal confidential task subjects or project information to the remote service. Raw conversations are not transmitted by this code; the confirmed disclosure is limited to the derived keywords and associated request metadata.

T09 · Insecure Skill Coding Practices

Error
Location
scripts/deep_skill_install.py
Finding

Windows absolute archive paths can escape into prefix-matching sibling directories

Content
View full analysis

Impact

The package can create or overwrite files in writable prefix-matching sibling directories, including another Skill's instructions or scripts. This can poison another Skill and lead to attacker-controlled behavior when that Skill is subsequently loaded or executed.

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
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (25)

Tainted flow: 'req' from os.environ.get (line 514, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · scripts/skill_feedback.py (reported line 516)May include surrounding context.

python
req = urllib.request.Request(url, data=body, headers=headers, method="POST")
    try:
        with urllib.request.urlopen(req, timeout=FEEDBACK_API_TIMEOUT) as resp:
            raw = resp.read()
            try:
                result = json.loads(raw)

Tainted flow: 'req' from os.environ.get (line 854, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · scripts/skill_feedback.py (reported line 857)May include surrounding context.

python
req = urllib.request.Request(url, data=body, headers=headers, method="POST")

    try:
        with urllib.request.urlopen(req, timeout=FEEDBACK_API_TIMEOUT) as resp:
            raw = resp.read()
            try:
                result = json.loads(raw)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied chunk primarily manages skill installation and removal, whereas the description advertises discovery, recommendations, and evaluation. Installation and recursive deletion are material state-changing capabilities, not merely implementation details of searching or scoring skills, and are not disclosed. Downloads also transmit a persistent X-Client-Id, a requestId loaded from previous search results, and optional agent-type information, enabling correlation not mentioned in the description. Registry access and credential use support the installer and are not, by themselves, evidence of malicious behavior. This chunk does not implement the advertised task-based search or trajectory-based evaluation; those could exist elsewhere, but would not resolve its undeclared installation and removal capabilities.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The core marketplace search and retrieval of examples or reviews fit the declared purpose. However, digest() implements a separate recurring-recommendation feature, and first_run_check() is invoked automatically after successful searches to write invitation markers and potentially emit first_run_invite. These are additional user-facing recommendation and invitation capabilities, not merely search implementation details. The code generates digest data but does not itself install a scheduler or send recurring messages. Credential loading, API access, result files, and version checks are supporting mechanisms rather than independent grounds for this finding. Trajectory-based evaluation is not demonstrated in this chunk, though other skill files could implement it.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied command explicitly requests installation of a specific skill. The description declares discovery, matching, and evaluation, but does not describe installation or explain why this fixed skill is relevant. This is a substantive additional operation rather than a demonstrated implementation detail of search or evaluation. The installer source is not provided, so its internal behavior and whether installation succeeds cannot be verified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

Installing a skill changes the agent's available capabilities and is materially distinct from the declared discovery and evaluation functions. The description does not disclose installation. The script implementation is not supplied, so its exact filesystem, network, or execution behavior cannot be established; the mismatch is based on the installation-oriented invocation shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

Trajectory scanning and feedback preparation align with the declared evaluation workflow; the absence of skill-search logic in this feedback-specific chunk is not itself a mismatch. However, the code additionally supports transmitting local media files and a persistent machine-level identifier to an external service, materially extending the described rating and trajectory-selection behavior. API credentials are obtained from ~/.meyo_agent/app.config.json or MEYO_API_KEY and sent as bearer authentication. Upload requires --confirmed, so these transfers are not automatic, but large attachments are uploaded before feedback validation and redaction checks. Attachment contents are not inspected or redacted. These are substantive undisclosed data-transfer and identification capabilities, rather than merely local feedback implementation details.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Skill ranking, trending-skill recommendations, and recommendation-history tracking reasonably support the declared discovery purpose. However, recommending standalone published works extends beyond finding or evaluating skills. The first-success invitation signal is also an additional, undeclared onboarding behavior, although the invitation's content and delivery are not shown. Because this chunk contains tests rather than the implementation, it establishes the behaviors being tested, not actual production network access or invitation delivery.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Trajectory inspection and ordinary feedback processing align with the declared purpose. However, the tested attachment workflow adds a material data-transfer capability: local media files can be read and transmitted to remote storage. The description mentions evaluating and giving feedback on skills but does not disclose media collection or upload. Confirmation and sensitive-text checks provide safeguards but do not resolve that disclosure gap. This conclusion concerns capabilities exercised by the tests; the supplied test file does not itself demonstrate live network transmission.

Content

No source excerpt is available for this finding.

Obtains recurring access to work context under privacy assurances that conflict with using derived keywords for remote recommendations.

High
Category
Prompt Injection
Confidence
98% confidence
Finding

The optional schedule repeatedly accesses recent conversations or workspace content while presenting the feature as local-only processing. Because the resulting keywords are used for remote recommendations, a one-time approval under inaccurate privacy assurances can lead to recurring disclosure of changing confidential work topics.

Content

No source excerpt is available for this finding.

Weekly recommendations promise no uploads without disclosing transmission of work-derived search keywords.

High
Category
Output Handling
Confidence
99% confidence
Finding

The invitation promises that work information is not uploaded, but the scheduled task uses work-derived keywords to request remote recommendations. Users therefore approve the feature without an accurate account of what leaves their environment.

Content

No source excerpt is available for this finding.

Large attachments are transmitted before the validation and redaction checks promised by the documentation.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The upload handler calls _process_attachments before validating or checking redaction of the feedback payload, and large attachments are immediately transmitted by that helper. Consequently, a subsequently rejected submission can already have disclosed screenshots or videos; attachment bytes also receive no content redaction, so textual sanitization does not protect secrets visible in the media.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Autonomous activation is permitted for broadly defined unfamiliar or unresolved tasks.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

The description permits autonomous activation for broadly defined unfamiliar or unresolved tasks, while the search workflow sends the user's task description to an external service. Without a disclosure or approval boundary, an agent may transmit confidential task context even when the user requested only local assistance.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 45)May include surrounding context.

若脚本不支持 --check-version,改用以下方式直接读取远程 SKILL.md frontmatter(GitHub 兜底):

bash
curl -s "https://api.github.com/repos/wheelry/deep-skill-finder/contents/SKILL.md" -H "Accept: application/vnd.github.v3+json" | {python} -c "import sys,json,base64,re; d=json.load(sys.stdin); c=base64.b64decode(d['content']).decode(); m=re.search(r'version:\s*\"([^\"]+)\"', c); print(m.group(1) if m else 'unknown')"
  1. 比较版本:
    • {current_version} < 远程最新版本 → 提示用户:

Reclassifies potentially untrusted package content as authoritative runtime instructions to overcome host non-execution boundaries.

Medium
Category
Prompt Injection
Confidence
96% confidence
Finding

The package explicitly targets hosts that treat its contents as attachment data or decline extra commands, then asserts that its own text is authoritative runtime procedure and should be executed anyway. This is a trust-boundary override attempt; the immediate requested action runs package-controlled Python and writes local state, even though the text characterizes it as harmless.

Content

No source excerpt is available for this finding.

The weekly recommendation workflow passes work-derived keywords to the search service despite promising that work context remains local.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The weekly workflow derives keywords from private conversations or workspace content and passes them to a remote recommendation service despite saying work content remains local and nothing is uploaded. Derived keywords can reveal confidential projects, customers, or work topics, making consent obtained under that assurance materially misleading.

Content

No source excerpt is available for this finding.

The skill mandates Chinese output without user language choice.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The mandatory rewriting rules explicitly require recommendation explanations to be written in colloquial Chinese. L177 similarly requires Chinese error messages, and neither requirement offers language choice or documents a locale-specific justification.

Content

No source excerpt is available for this finding.

The skill modifies the agent's installed skills despite advertising only skill discovery and evaluation.

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

The manifest describes finding suitable skills and evaluating previously used skills, but does not declare installation or removal. The implementation downloads and writes skill files into the agent's skills directory (L310–L327) and recursively deletes selected skill directories during uninstall (L374–L376), changing the agent's available capabilities rather than merely discovering or evaluating them.

Content

No source excerpt is available for this finding.

Search requests transmit a persistent installation identifier not disclosed by the task-based skill-finding manifest.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

Every semantic search sends an automatically generated, persistent installation identifier in X-Client-Id without a visible disclosure or opt-out. This enables longitudinal association of potentially sensitive task queries, and the feedback implementation uses the same identifier, allowing searches and feedback to be linked.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · scripts/deep_skill_search.py (reported line 667)May include surrounding context.

python
def _check_version_github(skill_name: str = "deep-skill-finder") -> str:
    """从 GitHub 主仓库 SKILL.md frontmatter 获取最新版本号。"""
    try:
        api_url = "https://api.github.com/repos/wheelry/deep-skill-finder/contents/SKILL.md"
        headers = {"Accept": "application/vnd.github.v3+json", "User-Agent": "deep-skill-finder/1.0"}
        token = os.environ.get("GITHUB_TOKEN")
        if token:

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/deep_skill_search.py (reported line 689)May include surrounding context.

python
"""从 SkillHub CLI 获取最新版本号(skillhub search 输出含 version 字段)。"""
    import subprocess
    try:
        result = subprocess.run(["skillhub", "search", skill_name], capture_output=True, text=True, timeout=15)
        output = result.stdout + result.stderr
        m = re.search(r'version:\s*(\d+\.\d+\.\d+)', output)
        if m:

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/deep_skill_search.py (reported line 703)May include surrounding context.

python
"""从 ClawHub CLI 获取最新版本号(openclaw skills info 输出含 version 字段)。"""
    import subprocess
    try:
        result = subprocess.run(["openclaw", "skills", "info", skill_name], capture_output=True, text=True, timeout=15)
        output = result.stdout + result.stderr
        m = re.search(r'version:\s*(\d+\.\d+\.\d+)', output)
        if m:

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · tests/test_skill_feedback.py (reported line 48)May include surrounding context.

python
"payload": {
                        "type": "custom_tool_call",
                        "input": {
                            "cmd": "sed -n '1,200p' /tmp/.codex/skills/demo-skill/SKILL.md"
                        },
                    },
                },

An ordinary search also triggers an unsolicited weekly-recommendation invitation and persists invitation state.

Low
Category
Excessive Agency
Confidence
93% confidence
Finding

Following a successful search, the code invokes first_run_check() and adds first_run_invite to the search output, explicitly intending the host Agent to display a weekly-recommendation configuration card. The helper at L521-L555 also writes an invitation marker and removes a legacy marker, extending a task-based search into an onboarding workflow for recurring recommendations that the manifest does not describe.

Content

No source excerpt is available for this finding.