Back to skill

Security audit

Office Prolonged Sitting & Posture Warning | 成人久坐/姿态预警(办公室)

Security checks for vulnerabilities and agentic risk

Overview

This skill needs review because it can upload workplace video despite an export-only selection and ignores the user's chosen API server.

Have the publisher fix export-only behavior and API-server selection before using sensitive footage. Confirm authorization for external processing, the provider's retention and deletion practices, and how stored accounts and tokens can be removed. Protect the workspace database and credential-bearing logs. The inspected artifacts support review for privacy and control failures, rather than a malicious verdict; remote-service protections and advertised real-time capabilities remain unverified.

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
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (29)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied code's primary purpose is general database and user-account management, not posture monitoring. It supports reading, writing, updating, and deleting user records containing names, email addresses, demographic data, and tokens, and initializes or alters the local database when a DAO is instantiated. Persistence could support a larger application, but this chunk shows no connection to posture measurements or reminders; its user-account functionality is materially different from the declared purpose. This assessment is limited to the supplied chunk and does not establish what other files implement. No camera access, network transmission, or token use is demonstrated here.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The visible implementation is generic skill scaffolding with a nonfunctional chat wrapper, materially different from the declared posture-monitoring behavior. No active subprocess execution is present. The imported base classes and services are not supplied, so their behavior cannot be assessed; this finding is limited to the provided code chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Ordinary formatting, polling, and filesystem helpers could support the declared skill and are not independently mismatches. However, the supplied chunk additionally implements an undeclared identity, account, authentication, and remote-service layer with local persistence and credential-bearing network access. It contains no camera processing, posture estimation, sitting-duration tracking, or posture-alert logic. These findings apply to the supplied chunk; they do not establish whether posture functionality exists elsewhere or whether these helper methods are actually invoked.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied code implements a general API-access layer rather than the declared posture-monitoring behavior. HTTP helpers could support a larger posture-analysis system, so their presence alone would not establish a mismatch. However, the storage-download and username-lookup capabilities have no demonstrated relationship to the declared purpose, and the chunk implements none of the stated monitoring functions. These are callable capabilities; the snippet does not establish that any requests execute automatically or that data is exfiltrated.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The visible implementation is an API-backed video-analysis and report-management client, not the declared real-time posture-monitoring system. Remote submission and polling could support posture analysis, but the supplied chunk does not establish that the backend performs the declared measurements or alerts. The batch-oriented workflow and historical-report retrieval materially differ from the stated behavior. Imported API implementations are not supplied, so their internal analysis capabilities cannot be verified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The history-listing feature is an explicit additional capability: it obtains the current Open ID and calls get_output_analysis_list. The main analysis path simply passes a file path or URL to get_output_analysis; no camera stream handling, sitting-duration state, posture measurements, or alert rules are visible. The imported implementation could perform posture-specific analysis, so its absence from this wrapper alone does not establish that such analysis never occurs. Nevertheless, the undeclared history access and materially different exposed workflow support a mismatch. Optional result-file output is a supporting implementation detail, not an independent mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The clearest mismatch is the additional identity-associated analysis-listing functionality: the code resolves an open ID and can call get_output_analysis_list instead of analyzing video. These are observable capabilities not disclosed by the description. The supplied chunk does not itself implement continuous camera monitoring, posture measurements, or alert thresholds; it delegates analysis to an unavailable module, so those advertised functions cannot be verified from this chunk alone. There is no visible basis to assert video uploading, external transmission, or malicious behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

The video-analysis commands are consistent with the declared purpose, and report formatting and saving are supporting implementation details. However, retrieving historical warning reports is an additional user-facing capability not disclosed by the real-time monitoring description. The snippet shows command examples rather than the module implementation, so it does not establish whether the stated measurements, timing logic, or alert thresholds are actually implemented.

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-office-worker-posture-warning-analysis"
description: "Using a fixed camera in the office (aimed at the workstation), the system analyzes office workers' sitting-posture video in real time, detecting continuous sitting duration, neck forward angle (head offset relative to shoulders), and back curvature (hunchback degree). | 通过办公区固定摄像头(对准工位)实时分析办公人员的坐姿视频,检测连续坐姿时间、颈部前倾角度(头部相对于肩部的偏移)、背部弯曲度(驼背程度)。当久坐时间超过预设阈值(默认1小时)且未起身活动时,输出'久坐提醒';当颈

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 an explicit warning that workplace footage is processed by an external API and reports are associated with cloud history.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

Although API calls, cloud history, and employee consent are mentioned elsewhere, the execution workflow does not explicitly disclose that supplied workplace footage leaves the local environment or establish authorization for that transfer. Users may authorize posture analysis without understanding the external processing and account-associated report history, exposing sensitive employee or bystander information.

Content

No source excerpt is available for this finding.

The documented --export-env-only mode does not prevent video analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The parser accepts --export-env-only and promises not to perform analysis, but the parsed flag is never checked. With an input supplied, execution still initializes identity and invokes analysis, violating a caller's explicit no-analysis expectation and potentially transmitting sensitive workplace footage.

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 content is submitted to the analysis API without disclosing the upload to the user.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The method reads the local file into a multipart payload and passes it to the analysis service without a disclosure or authorization check in this path. For office footage, an unexpected external transfer can expose employees, bystanders, and visible workplace information; file-format validation does not address that privacy boundary.

Content

No source excerpt is available for this finding.

The documented export-only option does not prevent video analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The export-only flag explicitly promises no analysis, yet execution never branches on its value. When combined with a video input, it still invokes analysis and can cause an unintended external submission despite the caller selecting a non-analysis mode.

Content

No source excerpt is available for this finding.

L071-L095 expose caller-directed HTTP mutation capabilities unrelated to posture monitoring.

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

The service exposes HTTP POST, PUT, and DELETE methods that forward caller-supplied URLs and arguments to RequestUtil. General-purpose remote modification, particularly DELETE, is not justified by the manifest's stated purpose of analyzing camera footage and producing sitting-duration and posture alerts; these methods are not restricted to a posture-analysis or notification API.

Content

No source excerpt is available for this finding.

The skill includes user-profile and credential persistence unrelated to posture monitoring.

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The User model stores email, birthday, demographic information, token, and open_token fields (L448–L465), and UserDao exposes inherited CRUD operations over these records (L470–L472). This exceeds the stated purpose of measuring sitting duration and posture and issuing reminders; additionally, DAO initialization automatically creates and migrates the sys_user table (L185–L215), rather than keeping user-account management separate.

Content

No source excerpt is available for this finding.

The ai_chat docstring claims to invoke the OpenClaw agent, but the implementation never executes it or produces an analysis result.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The docstring at L25–L34 explicitly promises a subprocess invocation of openclaw agent and an analysis result or session ID. However, the invocation is commented out at L61–L66, and the empty dictionary assigned at L68 causes result.stderr to raise an exception, which is caught without returning a result. This contradicts the method's stated analysis behavior rather than merely omitting an implementation detail.

Content

No source excerpt is available for this finding.

FileUtil.open silently truncates existing files without disclosing its destructive behavior.

Medium
Category
Output Handling
Confidence
95% confidence
Finding

Opening a path in 'w' mode immediately erases any existing contents, even before a caller writes data. This generically named helper has no docstring, comment, confirmation, or action log explaining that behavior; exception reporting does not disclose successful truncation.

Content

No source excerpt is available for this finding.

The request helper silently provisions remote user accounts and persists authentication tokens, capabilities unrelated to the declared posture-monitoring purpose.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

A routine HTTP call can automatically register a remote account using the current identity and cache the returned authentication tokens, without an explicit enrollment step in this flow. Authentication may legitimately support posture analysis, but implicit identity disclosure and persistent account linkage are privacy-sensitive in workplace monitoring.

Content

No source excerpt is available for this finding.

Ordinary HTTP requests can silently trigger remote account registration without disclosure.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The login payload explicitly requests silent registration, and the helper can invoke it automatically when credentials are unavailable. A caller expecting only a posture-service request can therefore trigger remote account creation and identity association without a separate enrollment decision.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
86% confidence
Finding

This POST transmits the selected identity as both openId and mobile to the configured health-service endpoint as part of implicit enrollment. External authentication is not inherently malicious, but the undisclosed transmission is a concrete privacy concern; the code does not demonstrate that this particular request uploads workplace video.

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 YAML files, including during module import.

Low
Category
Output Handling
Confidence
93% confidence
Finding

YamlUtil.load creates parent directories and writes a configuration file when the requested path is missing, without a warning, log, or explanatory comment/docstring. BaseEnum.init_subclass invokes this loader at L053 and L058, so merely defining the module's subclasses can cause filesystem writes that are unexpected for a configuration read.

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