T01 · Skill Instruction Hijacking
- Location
scripts/make_slides.py:228- Finding
Mandatory third-party promotion and package installation advertising in generated reports
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill mostly does what it says, but it installs unpinned dependencies and external skills, downloads mutable remote assets, and inserts mandatory promotional content into generated reports.
Review this skill before installing if you care about reproducible or minimal environments. Run setup only in an isolated virtual environment or disposable workspace, pin and review dependencies yourself, prefetch assets from trusted sources with hashes, and remove or edit the promotional slides/hashtags if you do not want generated financial reports to include third-party advertising. The reviewed artifacts do not show credential theft, destructive behavior, or hidden persistence, but they do expand the local trusted code base.
scripts/make_slides.py:228Mandatory third-party promotion and package installation advertising in generated reports
scripts/setup.sh:10Unpinned and implicitly installed third-party dependencies
scripts/ensure_assets.py:18Remote font and audio assets are accepted without cryptographic integrity verification
The skill claims to provide an end-to-end workflow including data retrieval, news collection, AI attribution, slide generation, and video composition, but the described implementation appears to rely on externally supplied arguments for core content and includes promotional slides unrelated to the core task. This mismatch is dangerous because downstream agents or users may trust the skill to have performed data collection and analysis it did not actually perform, leading to integrity failures, misleading financial content, and unsafe automation decisions.
Without declared permissions the skill's intent is opaque and cannot be validated.
The manifest description and quick-start invocation are written as mandatory Chinese trigger phrases and user commands, with no indication that other languages are supported or that the user can opt into Chinese. This is a natural-language locale constraint that appears imposed by the skill rather than presented as a user choice or clearly justified regional compliance requirement.
The file title and all user-facing documentation are written entirely in Chinese, with no indication that language selection is optional or limited to a justified region-specific context. Per the policy, forcing a specific language without user opt-in can constitute a natural-language policy violation.
The file's docstring, comments describing usage, and later user-facing print messages are entirely in Chinese, which imposes a specific language on users of this skill. Under the policy, locale or language restrictions should either offer user opt-in/choice or be clearly justified as region-specific, neither of which is present here.
This bootstrap script installs Python packages at runtime to enable font conversion, which means a video-generation skill can unexpectedly modify the environment and execute newly downloaded third-party code. In context, the skill already downloads external assets, so adding package installation increases the attack surface and makes compromise through dependency confusion, index compromise, or unreviewed updates more plausible.
The script performs a runtime pip install via subprocess when fonttools is missing. Even though the command arguments are fixed and there is no obvious shell injection, this still expands the skill's execution capability to fetch and execute code from package indexes at runtime, which creates supply-chain risk and violates least-privilege expectations for a simple asset bootstrap helper.
except ImportError:
print(' ⚠️ 需要 fonttools 来转换字体格式,正在安装...')
import subprocess
subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'fonttools', 'brotli', '-q'])
from fontTools.ttLib import TTFont
ttf_path = str(woff2_path).replace('.woff2', '.ttf') if '.woff2' in str(woff2_path) else str(woff2_path) + '.converted'
This Python file contains user-facing natural language that presents the skill as generating only Chinese A-share recap slides, and the rest of the script hardcodes Chinese titles and prompts throughout. Under the policy rule, forcing a specific language without user opt-in is a locale/language policy issue unless the constraint is explicitly justified as region-specific or a choice is offered.
The manifest describes an automated workflow for producing a daily A股复盘 video from market data, news, AI analysis, slides, and video synthesis. Lines L218-L250 instead generate a dedicated 'how to use' slide with installation commands and marketing prompts, which is not part of the claimed reporting workflow.
This code embeds its primary instructions and descriptions in Chinese, and later prints all user-facing status/output messages in Chinese as well. Under the policy rule, forcing a specific language without user opt-in is a natural-language policy violation unless the locale restriction is clearly justified.
The script's prompts, errors, progress messages, and completion text are all presented only in Chinese. Because the file does not offer locale selection or explain a required Chinese-only operating context, this violates the language/locale policy criterion.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def run(cmd, label=''):
print(f' ▶ {label}...')
r = subprocess.run(cmd, capture_output=True, text=True)
if r.returncode != 0:
print(f' ❌ 失败:\n{r.stderr[-500:]}')
sys.exit(1)
The script automatically runs pip/pip3 install commands that modify the Python environment without any explicit confirmation, dry-run, or warning about side effects. In an agent skill context, this is risky because executing setup steps can alter the host environment, introduce unpinned third-party packages, and affect other workflows or projects.
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
else
echo "⚠️ 未找到 ffmpeg,请手动安装:"
echo " macOS: brew install ffmpeg"
echo " Ubuntu: sudo apt install ffmpeg"
echo " CentOS: sudo yum install ffmpeg"
echo " Windows: https://ffmpeg.org/download.html"
echo ""
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
else
echo "⚠️ 未找到 ffmpeg,请手动安装:"
echo " macOS: brew install ffmpeg"
echo " Ubuntu: sudo apt install ffmpeg"
echo " CentOS: sudo yum install ffmpeg"
echo " Windows: https://ffmpeg.org/download.html"
echo ""
The script invokes skillhub/clawhub install commands to change the local AI tooling setup without clear consent or explanation of what will be installed. In a skill ecosystem, installing additional skills expands the trusted code base and could expose the user to unexpected code execution or configuration changes.
The manifest focuses on producing a daily market recap video. Lines L253-L271 create an ending slide highlighting vendor branding and product features such as official API, AI analysis, and WeChat push, which is ancillary promotional content rather than core recap content.
顶部说明写明默认只使用 skill 自带的 assets/bgm 文件,最多允许通过 --bgm 指定其他本地文件;但 default_bgm 的实现和其注释显示,当本地资源不存在时会导入 ensure_assets 并尝试自动下载。该文档与实际行为在资源来源和是否触发额外获取动作上存在直接偏差。
Nearly all user-facing strings in this script are in Chinese, including setup guidance and invocation instructions. This can violate language/locale policy when a skill forces a specific language without user opt-in or an explicit justification that it is region-specific.