T03 · Remote Payload Retrieval and Execution
- Location
SKILL.md:30- Finding
Remote Metadata Controls Mandatory Skill Installation and Updates
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill matches its stated FunAI creation purpose, but it asks agents to auto-update from mutable remote metadata and handles the user's API token in unsafe ways.
Review before installing. Only use this skill if you trust the publisher and the remote update endpoint, are comfortable storing a FunAI token locally, and can avoid raw absolute-URL API calls. Prefer disabling automatic skill updates, keeping the token out of project files or locking down config/.env permissions, and rotating the token if it may have been exposed.
SKILL.md:30Remote Metadata Controls Mandatory Skill Installation and Updates
scripts/api-client.sh:115FunAI Authorization Token Can Be Forwarded to an Arbitrary URL
scripts/api-client.sh:45Credential Configuration File Is Executed as Shell Code
CLAWHUB_PUBLISH_GUIDE.md:44Publishing Guide Installs an Unpinned Global npm Package
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.
把版本号和 changelog 改掉即可:
clawhub publish "/Users/lvke/Desktop/Funai-skill" --slug funai-skill --name "Funai-skill" --version 1.0.7 --changelog "Update skill instructions and resources" --tags latest
声明描述强调的是一个面向橙星梦工厂平台的对话式制作工具,核心能力包括创建项目、配置剧本、生成角色/分镜/视频、查询任务状态。实际代码则是一个离线批处理示例脚本,主要用于获取项目对应的 storyboard,汇总 scene 与 prompt 信息,并批量下载 scene 图片到本地目录以便审查。虽然代码确实与橙星梦工厂项目数据/API 相关,属于同一平台生态,但其主要功能并非声明中的“创作/生成/项目配置”,而是“分镜审查与图片导出”。此外,下载图片到本地文件系统是声明中未体现的能力。因此存在明显的描述与行为不一致。
The skill instructs the agent to fetch remote metadata and automatically install or update itself before performing operations. This is a supply-chain risk: a compromised remote endpoint or package URL could cause unreviewed code or instructions to be introduced into the agent workflow without informed user consent.
The instructions explicitly involve creating a .env file to hold an API token, which is direct credential material handling. Plaintext secret storage in project files is dangerous because secrets may be exposed to other tools, accidental commits, local users, or support/debug workflows.
cp config/.env.example config/.env
# 编辑 config/.env
# AIFUN_TOKEN=...
The skill shows the token value being placed into a configuration file field, reinforcing a workflow that persists sensitive credentials in plaintext. In an agent environment with file and shell capabilities, this materially raises the likelihood of unauthorized disclosure or reuse.
cp config/.env.example config/.env
# 编辑 config/.env
# AIFUN_TOKEN=...
Referenced artifact was not completely inspected
./scripts/api-client.sh setup-check
Referenced artifact was not completely inspected
./scripts/api-client.sh setup-check
Referenced artifact was not completely inspected
./scripts/api-client.sh setup-check
Referenced artifact was not completely inspected
./examples/create-comic.sh
Referenced artifact was not completely inspected
./examples/create-comic.sh
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
cp config/.env.example config/.env
# 编辑 config/.env,填入 AIFUN_TOKEN
bash scripts/api-client.sh setup-check
./scripts/api-client.sh check
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
cp config/.env.example config/.env
# 编辑 config/.env,填入 AIFUN_TOKEN
bash scripts/api-client.sh setup-check
./scripts/api-client.sh check
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
cp config/.env.example config/.env
# 编辑 config/.env,填入 AIFUN_TOKEN
bash scripts/api-client.sh setup-check
./scripts/api-client.sh check
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
cp config/.env.example config/.env
# 编辑 config/.env,填入 AIFUN_TOKEN
bash scripts/api-client.sh setup-check
./scripts/api-client.sh check
The README states that config/.env is directly sourced by a shell script. If an attacker can modify that file, any shell syntax, command substitution, or injected commands in .env will execute with the privileges of the user running the script, turning a credential file into a code execution surface.
scripts/api-client.sh 不是按 dotenv 语义做“纯文本解析”,而是直接:
source config/.env
因此:
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
SKILL_DIR="$(dirname "$SCRIPT_DIR")"
CONFIG_FILE="$SKILL_DIR/config/.env"
VERSION_FILE="$SKILL_DIR/VERSION"
BASE_URL="https://ai.fun.tv/service/workflow"
PROJECT_LINK_PREFIX="https://ai.fun.tv/#/comic/multiple?projectId="
The skill declares no explicit tool scope while its instructions clearly require network access, file reads, environment access, and shell execution. This creates an over-privileged, under-documented skill boundary, increasing the chance an agent will use sensitive capabilities without clear authorization constraints.
The skill directs the agent to collect a user token and write it into a local configuration file, but does not require secure storage, minimization, or a warning about credential handling. This increases the risk of token leakage through filesystem exposure, logs, backups, or later unintended file reads.
The skill content consistently directs agent behavior and user-facing outputs entirely in Chinese, with no indication that users may choose another language. Under the policy, forcing a specific language without user opt-in is a natural-language policy concern unless the locale restriction is explicitly justified.
The skill metadata and prompt content are written to drive Chinese-language interaction and fixed phrasing without checking the user's language preference. If auto-invoked for a user operating in another language, it can degrade informed consent, cause misunderstandings during step-by-step confirmations, and increase the chance of incorrect actions on an external content-generation platform.
The default prompt is a highly generic operational request to create an AI comic project on the FunAI platform, which can overlap with ordinary user intents about video/comic creation and API usage. This broad trigger scope can cause unintended skill invocation, leading the agent to apply this skill's prescriptive workflow, platform-specific actions, and defaults when the user did not explicitly request this integration.
Multiple user-visible strings, including error, usage, status, and recommendation messages, are written only in Chinese. This creates a language policy concern because the script does not provide language selection, fallback behavior, or a documented reason for restricting output to a single locale.
The script supports unattended execution via AUTO_CONFIRM, allowing it to bypass manual checkpoints and continue creating projects, submitting content, and initiating downstream processing. In this skill context, that increases the chance of unintended or unauthorized billable actions and reduces human review before state-changing API calls.
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
source "$SCRIPT_DIR/../scripts/api-client.sh"
AUTO_CONFIRM="${AUTO_CONFIRM:-}"
NOVEL_CONTENT='清晨,城市公园的林荫道上还带着尚未散尽的薄雾,年轻人林舟沿着青石板路缓缓慢跑。阳光穿过高大的树冠,在路面上投下细碎而温暖的光影,周围的一切都显得安静而清新。
At this checkpoint the script treats AUTO_CONFIRM=YES as approval to skip user review and proceed automatically. Because prior steps already submit project/script data and later steps trigger more processing, this weakens consent and operational safeguards around API-driven actions.
echo -e "${YELLOW}--- ${checkpoint_name}确认 ---${NC}"
echo "$modify_hint"
if [ "$AUTO_CONFIRM" = "YES" ]; then
echo "AUTO_CONFIRM=YES,自动确认继续。"
return 0
fi
This line implements automatic continuation when AUTO_CONFIRM is set, effectively nullifying the manual approval control. In the context of a media-generation platform, that can cause unwanted progression through content-generation stages without a human validating inputs or outputs.
echo "$modify_hint"
if [ "$AUTO_CONFIRM" = "YES" ]; then
echo "AUTO_CONFIRM=YES,自动确认继续。"
return 0
fi