Back to skill

Security audit

Visual Summarization Skill | 视觉摘要智述技能

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report account and credential management, report-history access, and other behavior beyond its stated image and video analysis purpose.

Consider avoiding installation until the extra account and report-history capabilities, local credential storage, external media processing, and behavior of the export-only option are clarified. The scan was partial because one local reference was missing, so the evidence is incomplete; the results do 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 (30)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied chunk contains no image/video processing or scene-description generation. Its substantive functionality is a user-account persistence subsystem, including sensitive token fields and database writes. Although a DAO can support a larger application, this chunk exposes user-management capabilities beyond ordinary media-analysis implementation details, with no visible connection to the declared purpose. No network transmission or token acquisition is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The media-analysis commands match the declared purpose, and detail selection and output saving are supporting features. However, retrieving historical reports is an additional capability not disclosed in the description, and the supplied content explicitly defines automatic triggers absent from the declaration. The underlying Python implementation is not provided, so further behavior cannot be verified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The main analysis flow is consistent with the declared purpose, and optional output-file writing is a supporting implementation detail. However, the independent history-listing mode accesses prior analysis records associated with an open ID without requiring input media. That is an additional, undeclared capability. The imported skill and identity utilities are not supplied, so their internal behavior and any underlying network or storage access cannot be verified from this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The main media-summary workflow matches the declared purpose. However, the independent --list workflow retrieves analysis records associated with a resolved user identity rather than analyzing supplied media. This is an additional, undeclared capability. Identity resolution and file output alone are supporting implementation details, not independent grounds for a mismatch. The supplied code does not expose the imported implementations, so their underlying storage, network access, and analysis behavior cannot be verified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

Generic formatting, polling, and API transport can support visual analysis and are not themselves mismatches. However, the supplied code also implements automatic account provisioning and persistent identity/credential management, capabilities not disclosed by the scene-description purpose. Its request wrapper also attaches configured credentials regardless of whether an absolute destination URL belongs to the configured service. These are material additional behaviors. The chunk contains utility definitions rather than a complete entry point, so their invocation by the overall skill and any actual credential disclosure cannot be established here.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

Media validation, submission, polling, and result formatting are supporting implementation details consistent with the declared purpose. However, get_output_analysis_list() independently calls page() and returns existing report records, adding report-image export URLs where applicable. Accessing historical reports is a distinct data-retrieval capability not described in the declaration. The supplied chunk does not establish the API's authentication behavior or the scope of records accessible.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

The AI-analysis endpoint configuration is consistent with supporting the declared purpose. However, DETAIL_EXPORT_URL references a health-service endpoint at /health/order/api/getReportDetailExport?id=, introducing a resource unrelated to the stated image/video analysis function. This is a configuration-level mismatch; the supplied chunk does not demonstrate actual access to or export of health reports.

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 skill requires Chinese output without user language opt-in.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The capability description explicitly requires combining the results into a fluent Chinese description. No language choice, user opt-in, or domain-specific justification accompanies this output-language constraint.

Content

No source excerpt is available for this finding.

Generic report-list phrases can trigger unrelated cloud history queries.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

Generic phrases such as report lists can automatically initiate cloud history retrieval without establishing that the user means visual-summary reports. Because the workflow requires direct retrieval and output using an automatically selected identity, an ambiguous request can cause unintended disclosure of historical analysis data.

Content

No source excerpt is available for this finding.

The analysis workflow lacks a clear privacy warning about external media processing.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The skill automatically processes local media through a cloud API without clearly explaining that the media itself will leave the device. Images and videos can contain sensitive personal or organizational information, so requesting a summary does not necessarily establish informed authorization for external disclosure.

Content

No source excerpt is available for this finding.

The documented export-only mode still performs media analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The parser advertises --export-env-only as not executing analysis, but the flag is never checked. With a valid input, execution still initializes identity and calls the analysis workflow, potentially submitting sensitive media when the caller explicitly selected a non-analysis mode.

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 the analysis API without disclosure that the file leaves the local system.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The implementation reads the complete local media file and passes its bytes to the analysis API without an explicit disclosure or consent gate in this path. Sensitive visual content can therefore be transferred externally unless authorization has already been established by the calling application.

Content

No source excerpt is available for this finding.

The skill can retrieve historical analysis reports, beyond describing supplied video or image content.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The manifest describes analyzing supplied videos or images to generate scene descriptions. However, get_output_analysis_list calls self.page(pageNum, pageSize, ...) at L158 to retrieve existing report records without any input media, providing a report-history access capability outside that stated purpose.

Content

No source excerpt is available for this finding.

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

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The --export-env-only flag explicitly promises not to run analysis, but its value is never checked before invoking the analysis workflow when an input is supplied. This violates an explicit no-execution boundary and could cause unintended processing of sensitive media or API usage, although the delegated implementation's exact external effects are not shown.

Content

No source excerpt is available for this finding.

The service includes username-based user lookup, which is unrelated to visual scene summarization.

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

At L098–L101, get_user_by_username delegates a caller-supplied username to a user-lookup utility. Looking up user accounts is a distinct capability not justified by the manifest's stated purpose of analyzing images or video and generating scene descriptions.

Content

No source excerpt is available for this finding.

The skill includes user-profile and credential database management unrelated to visual scene description.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The User model stores personal information and credentials, including real names, email addresses, birthdays, token, and open_token (L448-L465). UserDao exposes inherited read, update, and delete operations over these records, while DAO initialization creates and migrates the sys_user table (L180-L215); this account-management capability is not justified by the stated video/image scene-description purpose.

Content

No source excerpt is available for this finding.

The ai_chat docstring claims to invoke an AI agent, but the implementation never invokes it or returns analysis.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The docstring at L26 explicitly states that this method calls openclaw agent through subprocess. However, the subprocess call is commented out, and the replacement empty dictionary at L68 raises AttributeError when accessed as result.stderr; the exception is caught and the method returns None. This contradicts the documented AI-analysis operation rather than merely changing 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
97% confidence
Finding

Opening a path in mode 'w' immediately erases any existing contents, even before the caller writes data. The generically named helper has no docstring, comment, confirmation, or operation-specific log explaining this behavior; exception tracing does not disclose successful truncation.

Content

No source excerpt is available for this finding.

The request helper silently registers service accounts and persists user profiles and credentials, beyond the declared visual-summary function.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

An ordinary request can automatically invoke a login endpoint with registration enabled, then persist the returned user profile and authentication tokens through UserDao. This introduces remote account creation and credential retention beyond the declared visual-summary behavior without an explicit authorization step in this code; storage protections and any consent elsewhere cannot be verified from this file.

Content

No source excerpt is available for this finding.

HTTP authentication silently transmits the user's identifier and can register a remote account without disclosure.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The helper sends the selected identifier as both openId and mobile while enabling silent login and registration. This can disclose an identifier and create a remote account as a hidden side effect of requesting a visual summary; it overlaps with the account-provisioning issue rather than representing a separate exploit.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · 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 files, including during module import.

Low
Category
Output Handling
Confidence
90% confidence
Finding

YamlUtil.load creates parent directories and writes a YAML file when the requested configuration is absent, without a log, confirmation, docstring, or explanatory comment. BaseEnum.init_subclass invokes this loader while subclasses are defined (L044–L059), so merely importing the module can modify the filesystem without disclosing that side effect.

Content

No source excerpt is available for this finding.