Security audit

mdclaw多模态图片视频生成

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed remote MDClaw API client, so it is acceptable if users understand that their prompts, images, URLs, and account details go to an external service.

Install only if you trust MDClaw/appmiaoda with the prompts, URLs, images, and account credentials you provide. Use a dedicated API key and password, and avoid submitting confidential images, private URLs, or sensitive text unless external processing is acceptable.

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (4)

Context-Inappropriate Capability

Medium

Confidence: 84% confidence
Finding: The client exposes search, weather, and web summarization functions that are unrelated to the declared multimodal generation purpose, which increases hidden attack surface and data egress paths. In an agent skill ecosystem, undeclared or weakly justified capabilities can be invoked unexpectedly and send user inputs or URLs to a remote service without clear consent.

Missing User Warnings

Medium

Confidence: 97% confidence
Finding: The skill clearly routes prompts, uploaded images, and related request content to a remote third-party API gateway, but the documentation does not prominently warn users about this data transfer. This matters because users may submit sensitive text, images, or URLs under the assumption processing is local or first-party, leading to privacy, compliance, and confidentiality issues.

Missing User Warnings

Medium

Confidence: 96% confidence
Finding: The registration/login flow and general _call method send usernames, passwords, API keys, and arbitrary user queries to a remote gateway, but the client provides no explicit notice, consent prompt, or data-handling warning. In an agent context, this can lead to silent exfiltration of credentials and sensitive prompts to a third-party backend that users may not realize they are trusting.

Missing User Warnings

Medium

Confidence: 97% confidence
Finding: The upload_image method reads a local file and uploads its contents to a remote endpoint without an explicit warning that the file leaves the local system. For an agent skill, that is dangerous because local images may contain sensitive personal data, screenshots, documents, or metadata, and users may assume processing is local unless clearly told otherwise.

Static analysis

No suspicious patterns detected.