Back to skill

Security audit

Child Emotion Recognition (Crying/Tantrum/Low Mood) | 儿童情绪波动识别(哭闹/暴躁/低落)

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report account and credential handling, broad credentialed network requests, and other behaviors that exceed or do not clearly support its stated child emotion analysis purpose.

Treat the suspicious verdict as a reason to review carefully before installing. Consider whether you are comfortable sending child recordings to a remote service, whether that service and its data retention are acceptable to you, and whether shared account, local identity or token storage, and broad credentialed requests fit your setup. The skillspector scan was partial because a local reference was missing, so the evidence is incomplete.

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 (1)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/smyx_child_emotion_recognition_analysis.py
Finding

Default CLI initialization discards upstream sender identity and reuses a shared account

Content
View full analysis

Impact

In the stated shared-workspace deployment, one requester can receive another requester's child-emotion history and report links. Analysis submissions can also become associated with the same shared account. This is not a demonstrated server-side authentication bypass.

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
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (27)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The supplied chunk implements user-account persistence rather than multimodal emotion recognition. It contains no camera or microphone acquisition, audiovisual feature extraction, or emotion inference. Database utilities can support a larger application, but this chunk specifically manages identities, personal information, and tokens without showing a connection to the declared workflow. The mismatch is based on that materially different functionality, not merely on SQLite being an implementation detail. No network transmission or token exfiltration is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Generic formatting, polling, and HTTP helpers could support the declared feature and are not inherently mismatches. However, this chunk additionally implements an undeclared account-registration, identity-persistence, authentication, and billing workflow, accessing a local identity file and user database and communicating with configured remote services. These capabilities are not represented in the child-emotion-analysis description. No declared sensing or inference functionality is visible in the supplied chunk; other files could implement it, but that would not remove the undeclared identity and account behavior shown here. Most operations run only when the relevant methods are called.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The visible implementation does not substantiate the declared purpose. It contains generic scaffolding and a nonfunctional chat helper, not a real-time audiovisual emotion-analysis pipeline. No external command is actually executed. Imported base classes and services were not supplied, so their behavior cannot be assessed.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The supplied chunk contains no camera or microphone acquisition, facial-expression analysis, audio-feature extraction, motion analysis, or emotion classification. Its implemented purpose is generic backend API access, including account lookup and download-link retrieval that the description does not explain. Such a module could support a larger emotion-recognition system, but no connection to that purpose is demonstrated here. The methods only perform operations when invoked; the chunk does not establish automatic data collection, exfiltration, or malicious behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The visible workflow materially differs from the declared real-time monitoring capability: it reads an existing video file or forwards a video URL and waits for an API-generated report. Delegating inference to an API is not itself a mismatch, and report formatting and history retrieval are supporting features. However, the supplied code contains neither a live capture/stream-processing path nor evidence that the delegated analysis implements the advertised multimodal child-emotion recognition. The API implementation is not supplied, so its underlying inference capabilities remain unverified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The clearest undeclared capability is analysis-history retrieval: --list resolves an open ID and calls skill.get_output_analysis_list(open_id=...). The analysis path itself only forwards a video path or URL to skill.get_output_analysis; this chunk contains no camera capture, streaming loop, microphone processing, or explicit child-emotion recognition. The imported component may perform emotion analysis, so the absence of those algorithms here is not independently proof of a different analytical purpose. Nevertheless, the history-access capability and the visible batch-video interface materially exceed or differ from the stated monitoring workflow. Saving analysis output is a supporting implementation detail, not a separate mismatch.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · SKILL.md (reported line 3)May include surrounding context.

md
---
name: "smyx-child-emotion-recognition-analysis"
description: "Using fixed cameras (and optional microphones) at home or in kindergartens, AI multimodal analysis recognizes a child's facial expressions (eyebrow/eye shape, mouth-corner curvature), cry-sound features (pitch, frequency, duration), and body-motion amplitude (waving, stomping, curling up) in real time, and jointly identifies the child's typical emotional state: happy, sad, angry, fearful, etc. | 通过家庭或幼儿园内的固定摄像头(及可选麦克风),利用AI多模态分析技术实时分析儿童的面部表情(如眉眼形态、嘴角弧度)、哭声音频�

The request helper can attach shared service credentials to arbitrary caller-supplied HTTP destinations.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The helper accepts arbitrary absolute HTTP or HTTPS URLs and automatically attaches shared access tokens, an API key, and an authorization token without checking the destination. A caller able to influence the URL can send those credentials to an attacker-controlled server, while HTTP also exposes them in transit; compromised service access could affect sensitive child-related information.

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.

Instructions to upload recordings of children lack a privacy warning.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The input guidance explicitly instructs uploading audio-enabled video for crying analysis and recommends footage showing children's faces in private settings. Unlike the endpoint catalog, this is an upload instruction, but the document does not warn that identifiable images, voices, and private surroundings will be disclosed to the configured API service.

Content

No source excerpt is available for this finding.

The --export-env-only help text promises not to perform analysis, but the flag is ignored and analysis still runs.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The parser accepts --export-env-only and promises no analysis, but execution never checks the flag before identity initialization or analysis. When supplied with an input, a supposedly non-processing invocation can submit sensitive children's recordings for remote analysis and potentially incur charges.

Content

No source excerpt is available for this finding.

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · skills/smyx_analysis/scripts/skill.py (reported line 28)May include surrounding context.

python
result_json = JsonUtil.parse(result_json_pure_text, result_json_pure_text)

        result_json_common_ai_response = result_json.get("commonAiResponse") if isinstance(result_json,
                                                                                           dict) else result_json
        if result_json_common_ai_response:
            result_json = result_json_common_ai_response

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · skills/smyx_analysis/scripts/skill.py (reported line 33)May include surrounding context.

python
result_json = result_json_common_ai_response

        result_json_health_ai_response = result_json.get("healthAiResponse") if isinstance(result_json,
                                                                                           dict) else result_json
        if result_json_health_ai_response:
            result_json = result_json_health_ai_response

Local video contents are passed to an analysis API without disclosure that the file will be uploaded.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The local-file path reads the complete recording and passes it as a multipart file to the analysis API without a runtime disclosure or consent check for external transfer. Although the documentation mentions API use and general guardian consent, it does not clearly establish informed approval to transmit identifiable children's home or classroom recordings to the configured processor.

Content

No source excerpt is available for this finding.

The --export-env-only option promises not to perform analysis, but does not prevent it.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The --export-env-only flag is parsed but never consulted, so providing an input still reaches remote video analysis despite the explicit promise not to analyze. In this skill's context, that can unexpectedly disclose children's recordings and trigger billable processing.

Content

No source excerpt is available for this finding.

The skill includes persistent user-account and credential management unrelated to its stated emotion-recognition purpose.

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

The User model stores account identifiers, email, real name, and token/open_token credentials, and UserDao exposes inherited CRUD operations for these records (L448–L473). DAO initialization also creates and migrates the sys_user table (L180–L214), making this an implemented account-storage capability rather than merely unused field declarations. Managing persistent accounts and credentials is not justified by the manifest's camera/audio-based emotion-analysis scope.

Content

No source excerpt is available for this finding.

ai_chat claims to invoke an analysis agent, but its implementation never invokes one.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The docstring at L26 explicitly says this method calls the openclaw agent through subprocess, and L33-L34 promise an analysis result or session ID. However, the subprocess invocation is commented out; the code instead assigns an empty dictionary and accesses result.stderr, raising an exception that is caught and suppressed, so the method returns None without performing analysis.

Content

No source excerpt is available for this finding.

FileUtil.open silently truncates existing files without disclosing its destructive write mode.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

At L101, opening a path with mode 'w' immediately erases any existing contents. The generically named open helper has no docstring, explanatory comment, confirmation, or success-path log disclosing this destructive behavior; its exception handler only runs after a failure.

Content

No source excerpt is available for this finding.

HTTP requests can silently register a remote account and persist its credentials, beyond the described emotion-analysis workflow.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

An ordinary HTTP operation can trigger a separate remote login or registration with silent=1 and register=1, then persist returned credentials without an explicit enrollment step. This creates undisclosed account and identity-handling side effects in a child-focused workflow, where users need clear boundaries around external services and retained data.

Content

No source excerpt is available for this finding.

HTTP authentication silently transmits the current identity for remote login or registration.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The authentication fallback silently transmits the selected identity as both openId and mobile to a separate login or registration endpoint. The selected value can even be API_SECRET_KEY because it takes precedence when choosing the username, unnecessarily placing a credential in identity fields that may receive different logging and retention treatment.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
92% confidence
Finding

This POST is the transmission point for the implicit registration flow, sending the current identity and service-source information outside the local workspace. External requests are not inherently unsafe, but this one inherits the undisclosed registration and possible API-key-as-identity exposure described in the overlapping findings; the source does not show child audio or video being transmitted here.

Content

Scanner excerpt · skills/smyx_common/scripts/util.py (reported line 561)May include surrounding context.

python
"source": ConstantEnum.DEFAULT__SKILL_HUB_NAME
            }
            try:
                _response = requests.post(_url, json=_data)
                if _response.status_code == 200:
                    _response_json = _response.json()
                    if _response_json and _response_json.get("success"):

Configuration loading silently creates directories and writes missing configuration files.

Low
Category
Output Handling
Confidence
95% confidence
Finding

YamlUtil.load modifies the filesystem when the requested configuration file does not exist, without a log, confirmation, docstring, or explanatory comment disclosing that behavior. Because BaseEnum.init_subclass calls this loader at L053 and L058, merely importing configuration classes can create files unexpectedly.

Content

No source excerpt is available for this finding.

Dynamic attribute access via getattr()

Low
Category
Dangerous Code Execution
Confidence
50% confidence
Finding

Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.

Content

Scanner excerpt · skills/smyx_common/scripts/dao.py (reported line 243)May include surrounding context.

python
username = getattr(model, "username", None)
            if username:
                column_names = self.__model__.__table__.columns.keys()
                update_data = {key: getattr(model, key) for key in column_names if key != "username"}
                return self.update_by_username(username, **update_data)

            return None

Dynamic attribute access via getattr()

Low
Category
Dangerous Code Execution
Confidence
50% confidence
Finding

Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.

Content

Scanner excerpt · skills/smyx_common/scripts/dao.py (reported line 353)May include surrounding context.

python
column_names = self.__model__.__table__.columns.keys()

            for key in column_names:
                value = getattr(model, key)
                setattr(instance, key, value)

            session.commit()

Dynamic attribute access via getattr()

Low
Category
Dangerous Code Execution
Confidence
50% confidence
Finding

Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.

Content

Scanner excerpt · skills/smyx_common/scripts/dao.py (reported line 326)May include surrounding context.

python
if filters:
                for key, value in filters.items():
                    query = query.filter(getattr(self.__model__, key) == value)

            if offset:
                query = query.offset(offset)