Back to skill

Security audit

Large Model Visual Question Answering Skill | 大模型视觉问答技能

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners report behavior beyond its image question-answering description, including video analysis, history access, persistent account handling, and credential and data handling concerns.

The scanner recommendation is DO_NOT_INSTALL. Consider avoiding installation unless you have independently reviewed the skill and are comfortable with remote media processing, identity transmission, credential handling, persistent account data, and the reported mismatch between its description and capabilities. The scan was marked partial because a local path-like reference was missing, so the evidence is incomplete; a separate scanner completed without findings, which does not establish that the skill is safe.

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 (31)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The command-line interface explicitly targets MP4 videos and delegates inputs to skill.get_output_analysis. It has no parameter for asking a question about an image. The --list option separately invokes skill.get_output_analysis_list with a resolved OpenID, exposing an undeclared history-retrieval capability. Although the imported service implementation is not supplied, the visible interface and calls establish a materially different purpose from image VQA.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The supplied chunk implements user-account persistence rather than image understanding or question answering. It contains no image processing, vision-model calls, or natural-language answer generation. Although database utilities can support a larger application, this chunk manages personal details and tokens without any demonstrated connection to the declared VQA purpose. These capabilities are not represented in the description. No external transmission or malicious use of stored data is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied code implements an analysis-report workflow rather than the declared open-ended image VQA workflow. It explicitly requests a local video path or network video URL, reads local file contents for submission through analysis(), polls analysis_result(), and formats structured report results. It also exposes report-history retrieval through page() and constructs export links. File reading and service calls could support a legitimate VQA implementation, but the video-oriented inputs and report-management capabilities materially differ from the stated purpose. The supplied chunk does not demonstrate image-specific question answering; the underlying service implementation and supported file formats are not shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The primary image Q&A workflow matches the declared purpose. However, the independently accessible history-listing mode retrieves existing user-associated records rather than answering a question about a supplied image. This is an additional undeclared data-access capability. API delegation, response formatting, and optional file output are supporting implementation details, not separate mismatches. The supplied code does not establish unauthorized access or reveal the internals of the imported identity and API helpers.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

Generic utilities and API authentication can legitimately support a VQA skill, so their presence alone is not a mismatch. However, this chunk also implements automatic account provisioning and persistent identity management not disclosed in the description. Its request helper accepts absolute URLs and attaches configured credentials regardless of destination, extending credential access beyond a clearly bounded VQA backend. These behaviors are conditional on the relevant methods being invoked; the chunk does not establish that registration or credential transmission occurs on every skill execution.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

Generic AI-analysis endpoints could support visual Q&A, so those alone do not establish a mismatch. However, the explicit health-order report export configuration introduces a resource outside the declared purpose. This chunk only defines configuration; it does not itself make requests or export reports, and the behavior of imported components is not shown.

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: "visual-qa-analysis"
description: "Conducts open-ended Q&A on image content based on computer vision and large language models, supporting any questions to receive natural language responses. | 大模型视觉问答(VQA)技能,基于计算机视觉和大语言模型对图片内容进行开放式问答,支持任意提问得到自然语言回答"
version: "1.0.24"
license: "MIT-0"
---

# ❓ Large Model Visual Question Answering Skill | 大模型视觉问答技能
> **智能分析中枢** · 图片/视频智能分析 · 结构化报告 · 历史报告云端查询

---

## 🧭 技能概览 | Overview

| 模�

The request helper can disclose service credentials to arbitrary destinations, beyond the needs of visual question answering.

High
Category
Excessive Agency
Confidence
97% confidence
Finding

The helper accepts arbitrary absolute HTTP or HTTPS URLs and automatically attaches service tokens and the API secret without checking the destination's origin. Any caller that passes an attacker-controlled destination can therefore disclose credentials; plain HTTP also exposes them to network interception, neither of which is necessary for VQA.

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 an explicit privacy warning about sending image content and questions to the remote API.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow directs local image processing through a remote API without explicitly telling users that file contents and questions leave their environment. References to API usage provide some notice, but do not clearly establish informed authorization to transmit potentially sensitive documents or images to an external service.

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 visual Q&A.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The parser advertises --export-env-only as a non-analysis mode, but the flag is never checked and valid image/question arguments still reach the analysis call. Users relying on that promise can unintentionally transmit private images and questions; identity initialization also occurs before any mode-specific safeguard.

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

The declared image-Q&A skill handles remote inputs as video-analysis requests.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The manifest describes answering questions about images, but the implementation explicitly requests a local video path or video URL and places remote inputs in the videoUrl request parameter. This is a video-specific input contract rather than the image-input behavior advertised by the skill.

Content

No source excerpt is available for this finding.

Local video contents are submitted for API analysis without disclosure of the upload.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The implementation reads the complete local file and passes its contents to the analysis API, rather than performing local-only inspection. In the absence of a clear upload disclosure and authorization step, sensitive visual material can leave the user's environment unexpectedly; the actual accepted media types depend on configuration.

Content

No source excerpt is available for this finding.

The skill exposes identifier-associated analysis history retrieval beyond its stated image Q&A purpose.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The show_analyze_list function requests analysis history using an open_id. The --list branch at L077–L081 makes this capability independently accessible without an image or question, although account-associated history retrieval is not part of the manifest's stated purpose.

Content

No source excerpt is available for this finding.

The skill implements video analysis rather than the advertised image question-answering workflow.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The manifest promises natural-language answers to arbitrary questions about images, but the CLI accepts local or remote MP4 videos and provides no question argument. The analysis function at L031–L037 passes only the input path or URL to the underlying skill, so this entry point does not implement the advertised question-answering interaction.

Content

No source excerpt is available for this finding.

The documented export-only flag does not prevent analysis from executing.

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

The export-only flag is parsed but never checked, so supplying media with this flag still reaches the analysis call. This violates an explicit no-analysis expectation and can cause unintended processing, service charges, or media disclosure if the downstream implementation uploads the input.

Content

No source excerpt is available for this finding.

General-purpose HTTP PUT and DELETE capabilities are unrelated to visual question answering.

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

The manifest describes answering questions about images, but L077–L081 and L091–L095 implement HTTP PUT and DELETE operations against caller-supplied URLs. Network requests for image retrieval or model inference are expected, whereas general-purpose remote modification and deletion capabilities have no stated role in this skill. This file establishes the capabilities, not that they are invoked during a VQA request.

Content

No source excerpt is available for this finding.

The VQA skill includes persistent user-profile and credential management unrelated to image question answering.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The User model stores names, email addresses, demographic information, and token/open_token credentials (L448–L465), while UserDao exposes inherited database CRUD operations (L470–L473). DAO initialization also creates tables and migrates sys_user (L180–L215), implementing persistent account management rather than merely consuming credentials for a vision-model request. This capability is not justified by the manifest's image-Q&A purpose.

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

The generically named open helper uses mode 'w', which immediately erases an existing file's contents. No comment, docstring, confirmation, or user-facing message explains this destructive behavior; the exception handler only runs after a failure.

Content

No source excerpt is available for this finding.

Ordinary HTTP requests can silently register a remote account and persist returned credentials.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

When cached tokens are unavailable, an ordinary request can initiate silent registration or login and persist the returned profile and credentials. Hiding these identity and storage side effects inside a generic request helper prevents callers from reliably controlling account creation and credential retention for an image Q&A task.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
88% confidence
Finding

This POST sends the selected identity as both openId and mobile to a configured service while requesting silent registration. External networking is expected for hosted VQA, but implicitly transmitting an identity or potentially secret-derived value for account provisioning is a separate privacy-sensitive action not required merely to transport an analysis request.

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"):

Generic history keywords can trigger visual-Q&A history retrieval outside its intended scope.

Low
Category
Trigger Abuse
Confidence
93% confidence
Finding

The examples include generic history keywords and instruct automatic cloud-history retrieval, which can cause unrelated requests about history to expose visual-analysis records in the conversation. Automatic reuse of a local default identity increases the privacy risk on shared agent installations.

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
92% confidence
Finding

YamlUtil.load creates parent directories and writes a YAML file when the requested configuration is absent, without a log, prompt, comment, or docstring disclosing this side effect. BaseEnum.init_subclass invokes this loader during class definition (L052–L058), so importing the module can modify the filesystem even when the caller only intends to read configuration.

Content

No source excerpt is available for this finding.