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

The skill is rated suspicious because scanners report credential handling, persistent account data, and sensitive media processing that is not clearly disclosed or limited to its stated emotion-analysis purpose.

Consider avoiding installation unless you can verify the media upload destination and consent process, restrict where credentials may be sent, and understand why the skill registers accounts and stores identity and token data. The scan was partial because a local reference was missing, so some evidence may be incomplete; the scanners did not establish malicious intent.

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 (25)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The visible code does not implement the declared emotion-recognition purpose. Its agent execution code is disabled, so it should not be characterized as actually launching commands. Behavior inherited from the unavailable BaseSkill and ApiService implementations cannot be assessed; the mismatch finding is limited to the supplied chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Generic formatting, polling, and HTTP helpers can legitimately support an emotion-recognition skill and are not themselves mismatches. However, this chunk also implements an undisclosed account lifecycle, persistent identity/credential handling, and a paid-service workflow beyond the declared visual-analysis purpose. Its HTTP wrapper attaches credentials even to explicitly supplied absolute URLs, with no visible host allowlist. No facial processing, emotion inference, intensity scoring, or abnormal-emotion marking appears in the supplied chunk; those features could exist elsewhere, but that does not explain the undisclosed account and billing behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied chunk implements general user-account persistence rather than facial-image processing or emotion inference. Database support alone could be an implementation detail, but this code specifically manages identity, demographic, and token records without any visible connection to emotion results. That constitutes an undeclared capability in the supplied code. No camera access, emotion scoring, abnormal-emotion detection, or network transmission is shown. This assessment is limited to the provided chunk; other files could implement the advertised recognition features.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

The supplied chunk exposes broad API and account-lookup functionality rather than the declared emotion-recognition behavior. HTTP wrappers alone could be supporting infrastructure and would not establish a mismatch, but the additional user-lookup and generic resource-management capabilities are not connected to the declared purpose here. These methods execute only when called; the snippet does not establish actual data transmission, destination services, or malicious activity. Other, unseen files could provide context linking some of this infrastructure to the skill.

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: "1.0.24"
license: "MI

The face-analysis workflow lacks a user-facing privacy warning about external API processing of sensitive media.

High
Category
Output Handling
Confidence
92% confidence
Finding

Although API use is mentioned, the workflow does not clearly tell users that their local face media will be transmitted to an external processor before analysis. Facial imagery and linked mental-health inferences are particularly sensitive, making this omission consequential for informed authorization.

Content

No source excerpt is available for this finding.

The HTTP helper can send service credentials to arbitrary destinations, beyond the emotion-recognition service's needs.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The helper accepts arbitrary absolute HTTP or HTTPS URLs and automatically attaches service tokens and an API key without checking the destination. An attacker able to influence a caller's URL can receive those credentials, while HTTP destinations additionally expose them in transit; emotion-recognition functionality does not require unrestricted credential forwarding.

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 documented export-only mode still performs emotion analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The export-only flag is parsed but never checked, so supplying valid media still reaches analysis despite the explicit promise not to analyze. A user relying on that promise could unintentionally transmit sensitive facial media for remote processing.

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 API without disclosure of the upload.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The code reads the complete local media file and passes it as a multipart upload to the analysis API without an upload disclosure or authorization check at this boundary. In a facial-emotion workflow, this can disclose identifiable imagery and material used to generate sensitive psychological inferences.

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 (L68-L69, L88-L95).

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The export-only option is accepted but execution never branches on its value, allowing analysis to proceed when media arguments are supplied. This violates the explicit no-analysis promise and can cause unintended external processing of private video.

Content

No source excerpt is available for this finding.

The skill includes persistent account and token management unrelated to its declared emotion-recognition purpose.

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

The User model stores identity information and token/open_token fields (L448–L465), and UserDao exposes the inherited database CRUD operations for these records (L470–L473). Persistent account and credential management is not an obvious requirement of visual emotion analysis or abnormal-emotion marking, and the manifest does not declare an account-management or authentication capability.

Content

No source excerpt is available for this finding.

ai_chat documents agent execution and an analysis result, but performs no execution and always returns None.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The docstring at L26–L34 states that this method invokes openclaw agent through subprocess and returns an analysis result or session ID. However, command execution is commented out at L61–L66, and the empty dictionary assigned at L68 causes result.stderr at L70 to raise an AttributeError, which is caught and suppressed at L81–L82. Consequently, this method cannot perform the documented analysis and silently returns None.

Content

No source excerpt is available for this finding.

The generic file-opening helper silently truncates existing files.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

FileUtil.open opens every supplied path in 'w' mode, immediately erasing existing contents. Its generic name does not disclose this destructive behavior, and there is no explanatory comment, docstring, confirmation, or success-path log; exception tracing only occurs after a failure.

Content

No source excerpt is available for this finding.

Ordinary HTTP requests can silently trigger remote account registration.

Medium
Category
Output Handling
Confidence
95% confidence
Finding

A routine request can automatically invoke a remote login endpoint with both silent and register enabled when tokens are missing. This creates externally persistent identity state and discloses identifiers without an explicit enrollment step, which is particularly concerning for a skill advertised for mental-health-related use.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
96% confidence
Finding

External transmission is not inherently unsafe, but this registration payload uses the selected username as both openId and mobile. The caller preferentially selects API_SECRET_KEY as that username, so when tokens are missing a configured API secret can be transmitted as ordinary account-profile data, potentially exposing it to profile storage and logging.

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”) is designated as an automatic trigger without an emotion-analysis scope constraint.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

The examples designate the generic phrase 'historical reports' as an automatic trigger, while other instructions require immediate cloud retrieval and presentation. This can cause unrelated requests to retrieve and expose sensitive emotion-report history without sufficiently establishing the user's intended scope.

Content

No source excerpt is available for this finding.

Configuration loading silently creates directories and files without disclosing the write side effect.

Low
Category
Output Handling
Confidence
94% confidence
Finding

YamlUtil.load creates parent directories and writes a YAML file when the requested configuration does not exist, without a warning, log, docstring, or explanatory comment. BaseEnum invokes this loader during subclass initialization (L052–L058), so importing the module can modify the filesystem even though the operation is presented as configuration loading.

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 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:

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/util.py (reported line 30)May include surrounding context.

python
elif isinstance(cur, dict):
                        cur = cur.get(part, default)
                    else:
                        cur = getattr(cur, part)
                return cur
            except Exception:
                return default