Back to skill

Security audit

Sora2数据助手

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this Sora2 skill suspicious because scanners report misleading consent records and other mismatches between its described purpose and behavior, alongside privacy and installation concerns.

Consider avoiding installation until the consent-record behavior and download-denial handling are clarified, privacy disclosures are added, and the installation package is pinned to a reviewed version. The scan is partial: although all 18 bundled components were scanned, some local references could not be resolved, so the evidence is incomplete.

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
  • Rogue AgentSelf-Modification, Session Persistence
Findings (22)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Local compliance bookkeeping could support the declared workflow, but this code does more than maintain implementation metadata: it asserts that confirmation occurred without any corresponding interaction or check. This is an undeclared consent-state modification that could cause downstream logic to treat confirmation as already satisfied. The mismatch is not merely that this chunk lacks API operations.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
89% confidence
Finding

The chunk does not perform the declared API or creative-workflow operations. Its substantive action is recording a completed legal review, which is not disclosed in the description. Although ordinary operational logging can be a supporting implementation detail, asserting completion of a legal review is a materially different compliance action. No network transmission is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

Local state and audit logging could be supporting implementation details. The material discrepancy is that this code asserts successful confirmation unconditionally, rather than checking or recording a demonstrated confirmation step. That behavior is not disclosed by the declared creative-workflow purpose and conflicts with its confirmation-gated framing. No API requests, media operations, or external data transfers occur in this chunk.

Content

No source excerpt is available for this finding.

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · SKILL.md (reported line 5)May include surrounding context.

md
slug: sora2
displayName: Sora2 数据助手 | 动态帖子 / 评论 / 二创 / 用户
name: sora2
description: Sora2 内容创作与作品分析 skill,通过 MaxHub API 查询作品、用户、评论、回复、二创/Remix、下载/媒体信息,并支持上传图片、创建视频等创作相关写操作。适合 AI 视频创作工作流、作品复盘、二创传播分析和素材管理。属于 read_write skill;agent 对 create/upload/download 等非只读能力必须按 confirmation 规则确认,普通查询优先走 recipes 与字段流字典。所有请求发送到 https://www.aconfig.cn。
license: MIT-0
metadata:
  author: maxhub

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · SKILL.md (reported line 26)May include surrounding context.

md
- curl
    env:
    - name: MAXHUB_API_KEY
      description: Sora2 内容创作与作品分析 skill,通过 MaxHub API 查询作品、用户、评论、回复、二创/Remix、下载/媒体信息,并支持上传图片、创建视频等创作相关写操作。适合 AI 视频创作工作流、作品复盘、二创传播分析和素材管理。属于 read_write skill;agent 对 create/upload/download 等非只读能力必须按 confirmation 规则确认,普通查询优先走 recipes 与字段流字典。所有请求发送到 https://www.aconfig.cn。
      required: true
      sensitive: true
    network:

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · SKILL.md (reported line 32)May include surrounding context.

md
network:
    - https://www.aconfig.cn
    riskLevel: medium
    defaultMode: recipes_first_confirm_write
    skillClass: maxhub-api-skill
    platform: sora2
    authType: bearer_env

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · SKILL.md (reported line 94)May include surrounding context.

md
## 1. 简介

Sora2 内容创作与作品分析 skill,通过 MaxHub API 查询作品、用户、评论、回复、二创/Remix、下载/媒体信息,并支持上传图片、创建视频等创作相关写操作。适合 AI 视频创作工作流、作品复盘、二创传播分析和素材管理。属于 read_write skill;agent 对 create/upload/download 等非只读能力必须按 confirmation 规则确认,普通查询优先走 recipes 与字段流字典。所有请求发送到 https://www.aconfig.cn。

> **架构亮点**:本 Skill 采用三层结构 — **Atomic 原子层**(端点的标准化封装)+ **Recipe 编排层**(多步业务场景脚本)+ **Reference 详情层**(端点 5 区契约)。Agent 可按"目标 → recipes 匹配 → atoms 映射 → reference 详情"四步路径精准调用,杜绝臆造。

The confirmation gate requires Chinese-language consent and explicitly rejects English consent without offering a language choice.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The accepted confirmations are exclusively Chinese, while the explicit English affirmative 'yes' is classified as ambiguous and rejected. This imposes a language requirement without user opt-in; the compliance context does not explain why consent must be expressed in Chinese.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skillhub'.

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The installation command invokes an unpinned npm executable through npx, potentially downloading and executing a changing package with the user's privileges. Although this is a CLI rather than an MCP server, a compromised future package release could expose local files and credentials during installation.

Content

No source excerpt is available for this finding.

The image-to-video example lacks a privacy warning that the user's image is sent to external services through MaxHub.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The example invites users to generate a video from their image, using the image-upload capability described on line 26. Although line 3 identifies MaxHub as the API intermediary, the README does not explain that user-provided images are transmitted to MaxHub and the generation service, leaving the privacy impact undisclosed.

Content

No source excerpt is available for this finding.

The compliance gate requires Chinese-language consent without offering a language choice.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The acceptance rules allow only Chinese confirmation phrases and explicitly reject English responses such as “yes.” This forces users to provide consent in Chinese to proceed, without a language opt-in or a documented justification for the restriction.

Content

No source excerpt is available for this finding.

The retrieval workflow sends user-provided Sora URLs or IDs to a third-party API without a privacy disclosure.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The instructions direct the agent to use a user-provided Sora URL or post ID in API requests to www.aconfig.cn, the service identified at L003. Although the API destination is documented, there is no warning that these user-supplied identifiers will be shared with a third party rather than sent directly to Sora.

Content

No source excerpt is available for this finding.

Treats paid or restricted download denials as a reason to offer an alternate video URL.

Medium
Category
Prompt Injection
Confidence
94% confidence
Finding

The endpoint-specific instructions explicitly offer an alternate video URL when downloading is denied because content is paid or restricted. This treats an authorization or entitlement decision as an availability failure and could facilitate unauthorized downloading if the playback URL remains accessible; actual bypass depends on upstream enforcement.

Content

No source excerpt is available for this finding.

The image-to-video workflow uploads user-provided image data without a privacy disclosure.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The executable workflow forwards the supplied image URL for upload without a media-specific disclosure that the image will be retrieved and processed by external services. Generic write confirmation does not clearly convey this privacy consequence, so users could submit private images or signed access URLs without understanding the exposure.

Content

No source excerpt is available for this finding.

The lookup workflows lack a user-facing warning that usernames and user IDs are sent to a third-party API.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

These chain recipes direct API calls using supplied usernames and user IDs, rather than merely listing available endpoints. L003 identifies https://www.aconfig.cn as the destination, but the file does not instruct the agent to disclose that these lookup inputs will be transmitted to an external service.

Content

No source excerpt is available for this finding.

L151–L156 allow querying the MaxHub account balance, a capability outside the declared Sora2 content-creation and analysis scope.

Low
Category
Excessive Agency
Confidence
87% confidence
Finding

The whitelist authorizes GET /api/maxhub/balance, which exposes account-level balance information rather than Sora2 content, users, media, or creation tasks. The manifest does not declare account-balance management or a credit-checking workflow that would justify this additional capability.

Content

No source excerpt is available for this finding.

The out-of-scope response mandates a Chinese-language reply without user language opt-in.

Low
Category
Output Handling
Confidence
92% confidence
Finding

The instruction tells the agent to reply directly with a fixed Chinese sentence whenever an unsupported capability is requested. It provides no language choice or instruction to adapt that response to the user's language, and the skill's purpose does not justify a Chinese-only response.

Content

No source excerpt is available for this finding.

The trigger table retains broad single-word triggers despite the explicit claim that all such triggers were removed.

Low
Category
Excessive Agency
Confidence
96% confidence
Finding

L010 states that every recipe's trigger_keywords has been narrowed from broad single words to complete intent phrases. However, L023 retains standalone “二创” and “Remix”, and L029 retains “Cameo” and “出镜”; these identify topics rather than explicit requests to run the associated queries. This contradicts the documented routing safeguard, although the separate requirement to clarify ambiguous intent limits the impact.

Content

No source excerpt is available for this finding.

The trigger “谁最火” (“who is most popular?”) does not specify the Cameo domain.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

This phrase can refer to popularity in many unrelated contexts, but it activates a Cameo leaderboard and profile lookup. Unlike the neighboring Cameo-specific trigger, it has no explicit contextual restriction or exclusion condition.

Content

No source excerpt is available for this finding.

Standalone triggers such as “Cameo” and “出镜” are too broad for a user-appearance lookup recipe.

Low
Category
Trigger Abuse
Confidence
92% confidence
Finding

“出镜” (appearing on camera) and “Cameo” can occur in general discussions without requesting a lookup of a user's appearance records. Although the recipe requires a user_id, its activation description does not explicitly require lookup intent and an identified target user.

Content

No source excerpt is available for this finding.

Batch-operation triggers include everyday words without requiring a data-collection context.

Low
Category
Trigger Abuse
Confidence
95% confidence
Finding

The keyword list includes “全部” (all), “所有” (every), and “循环” (loop), which can appear in ordinary requests unrelated to bulk data collection. Although the description names bulk collection, the matching rules provide no contextual qualification or exclusions to prevent unrelated requests from triggering that warning.

Content

No source excerpt is available for this finding.

The confirmation rules require Chinese-language consent without offering a language choice.

Low
Category
Output Handling
Confidence
97% confidence
Finding

Every accepted confirmation is Chinese, while English “yes” is explicitly rejected and the fallback prompt directs the user to enter “同意”. This imposes a language-specific interaction requirement without user opt-in or a documented justification; the issue is the enforced confirmation language, not merely that the file is written in Chinese.

Content

No source excerpt is available for this finding.