Back to skill

Security audit

Visual Emotion Recognition Skill | 人体视觉情绪识别技能

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report behavior beyond its stated emotion-analysis purpose, including credential handling, persistent user data, and privacy and execution-boundary concerns.

Consider the suspicious verdict before installing. In particular, review the reported remote video processing, credential use with unrestricted URLs, automatic account handling, and storage of user data and tokens. The scan is partial because one local reference was missing, so it cannot establish that the artifact is safe or fully account for that reference.

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
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (27)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied chunk contains no face-image processing, visual-model inference, emotion classification, intensity scoring, or abnormal-emotion marking. Its implemented purpose is database and user-record management, including fields such as real name, email, birthday, token, and open_token. Database access can support a larger application, but this chunk does not establish a connection to emotion recognition, and its user/token-management scope is materially different from the declared description. This assessment is limited to the supplied code; it does not establish whether emotion-recognition functionality exists elsewhere. No network transmission or token collection is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Formatting, polling, and ordinary API helpers could support the declared recognition feature. However, automatic account provisioning, credential-file access, persistent identity storage, and credential-bearing requests are substantial behaviors not disclosed by the description. The request helper also accepts arbitrary absolute URLs without restricting credential attachment to the configured service. No facial-image processing, emotion inference, intensity quantification, or abnormal-emotion marking appears in this chunk. Because this is a shared utility module, that absence does not establish that the full skill lacks those features; the mismatch rests on the additional identity and credential-handling behavior shown here.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The supplied chunk does not substantiate the declared visual emotion-recognition purpose. It contains generic scaffolding and a disabled chat integration, not a functioning facial-emotion pipeline. The commented subprocess code does not execute and is not evidence of active command execution. Imported base and API implementations were not supplied, so they could contain additional functionality; this assessment is limited to the visible code.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied chunk implements a generic API-access layer rather than facial analysis or emotion recognition. User lookup and storage-download URL retrieval are not connected to the declared emotion-monitoring workflow in this code. Although HTTP helpers could support a larger emotion-recognition implementation, no such integration is shown here. This assessment is limited to the supplied chunk; it does not establish malicious behavior or reveal the concrete endpoints, authentication, or data handled by the imported utilities.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

Video input handling, delegated analysis, and result export are plausible supporting features of the declared emotion-analysis purpose. However, resolving a user identity and retrieving that user's analysis history is a separate data-access capability absent from the description. The supplied chunk does not expose the underlying analysis implementation, so real-time emotion recognition, intensity quantification, and abnormal-emotion marking cannot be verified; their absence from this wrapper alone does not establish a 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: "human-emotion-recognition-analysis"
description: "Uses visual AI on frontal faces to recognize multi-dimensional emotions like happiness, sadness, depression, calmness, anger, surprise, and fear in real-time. Supports emotion intensity quantification and abnormal emotion marking, suitable for human-computer interaction and mental health monitoring. | 人体视觉情绪识别技能,基于正面人脸视觉AI实时识别快乐、悲伤、抑郁、平静、愤怒、惊讶、恐惧等多维度情绪状态,支持情绪强度量化与异常情绪标记,适配人机交互、心理健康监测场景"
version: "1.0.25"
license: "MI

The generic request helper attaches service credentials to unrestricted absolute URLs.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The helper accepts arbitrary absolute HTTP or HTTPS URLs and then automatically adds X-Access-Token, X-Api-Key, and Authorization credentials. If a caller supplies an attacker-controlled URL, those credentials are sent to that host; allowing HTTP additionally exposes them to network interception.

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.

The analysis workflow lacks a privacy warning about remote processing of facial media and cloud-accessible emotion reports.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow directs automatic processing of local facial media without a clear pre-upload privacy notice or consent step. Although API use is mentioned elsewhere, users are not clearly informed about transferring identifiable faces and potentially sensitive mental-health-related inferences to a remote service.

Content

No source excerpt is available for this finding.

The --export-env-only option promises not to perform analysis, but the code ignores it and proceeds with the normal analysis workflow.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The parser accepts --export-env-only and promises not to analyze, but the parsed flag is never checked. With media arguments present, execution still initializes identity and invokes analysis, potentially transmitting sensitive facial media despite an explicit non-analysis request.

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 submitted to an analysis API without disclosing the upload.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The implementation reads the entire local media file and passes its bytes to the remote analysis operation without an explicit consent gate. In a facial-emotion workflow, this crosses a significant privacy boundary because identifiable imagery and sensitive inferred attributes can leave the local environment.

Content

No source excerpt is available for this finding.

The --export-env-only option promises not to perform analysis, but execution still invokes video analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The export-only flag is parsed but never used to alter execution, so supplying it with a media input still invokes video analysis. This violates the advertised no-analysis boundary and can cause unintended remote processing of private recordings.

Content

No source excerpt is available for this finding.

The skill includes general-purpose user and credential persistence unrelated to its declared emotion-analysis purpose.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The model stores token and open_token as ordinary strings in a workspace-shared SQLite database, with no application-level encryption or explicit restrictive file permissions shown. A party able to read that database or its backups could recover service credentials; the purpose mismatch alone is not the vulnerability, but retaining credentials alongside identifiable user data increases exposure.

Content

No source excerpt is available for this finding.

FileUtil.open silently truncates existing files without disclosing its write-only behavior.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The generically named open method always uses mode 'w', which immediately erases an existing file's contents. There is no confirmation, user-facing log, docstring, or comment explaining this destructive behavior; exception tracing does not disclose successful truncation.

Content

No source excerpt is available for this finding.

The request helper silently provisions remote user accounts and persists their credentials, beyond the declared emotion-analysis purpose.

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

An ordinary request can implicitly register or log in a remote user, transmit the selected identity, and persist returned credentials without an explicit enrollment step. Identity selection can use API_SECRET_KEY, so the registration payload can also repurpose a secret as mobile and openId; this is an unexpected privacy and credential-handling side effect for an emotion-analysis skill.

Content

No source excerpt is available for this finding.

HTTP requests can silently initiate remote user registration and transmit the current identity.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The hidden authentication request sets both silent and register to 1 and sends the selected identity as openId and mobile. This makes remote account creation and identity disclosure side effects of a generic HTTP operation rather than explicit user-approved actions, which is particularly concerning in a mental-health-related workflow.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
92% confidence
Finding

External transmission alone is not a vulnerability, but this request can send API_SECRET_KEY as both openId and mobile because that secret takes precedence when selecting the username. The automatic registration path therefore risks placing a reusable secret into identity fields and downstream account records or logs, without an explicit enrollment decision.

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"):

The generic phrase “历史报告” (“historical reports”) automatically triggers emotion-report retrieval without requiring emotion-recognition context.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

The examples include the generic phrase 'historical reports' as an automatic trigger, while the rules require immediate cloud retrieval and display. An unrelated report request can therefore cause unintended access to and disclosure of sensitive emotion-analysis history, particularly on shared interfaces.

Content

No source excerpt is available for this finding.

Configuration loading silently creates directories and files, including during module import.

Low
Category
Output Handling
Confidence
90% confidence
Finding

YamlUtil.load creates parent directories and writes a configuration file when the requested path does not exist, without a log, prompt, docstring, or explanatory comment disclosing this side effect. BaseEnum.init_subclass invokes this loader when configuration classes are defined (L044–L058), so merely importing the module can modify the filesystem.

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)

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 441)May include surrounding context.

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

            return query.scalar()
        finally: