T07 · Tool Hijacking and Spoofing
- Location
scripts/bee.py:228- Finding
Execution of an Unverified Python Script Outside the Audited Skill Directory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is mostly aligned with video publishing automation, but it uses broad external publishing/storage authority and an unaudited sibling script path that users should review before installing.
Install only if you trust the publisher, control the configured OSS bucket, understand that media and metadata may leave your machine, and are comfortable with the fixed Feishu table destination. Review or remove the Yixiaoer external script dependency before using publishing, and prefer --no-feishu or --draft until destinations and permissions are confirmed.
scripts/bee.py:228Execution of an Unverified Python Script Outside the Audited Skill Directory
scripts/bee.py:39Operational Metadata Is Directed to a Hardcoded Feishu Table
The description promotes one-step automation for downloading, uploading, multi-platform distribution, and record creation but does not warn users that external content may be transmitted and published using stored credentials. In this context, the missing warning is especially dangerous because the skill performs multiple irreversible or externally visible actions across OSS, publishing platforms, and Feishu.
The skill declares powerful capabilities in practice—environment-secret access, file access, network operations, and shell execution—without any explicit tool-scope or permission boundaries in the manifest. In an agent environment, this increases the chance of overbroad invocation and unintended execution against local files, secrets, or external services using the user's credentials.
The natural-language triggers are broad enough to match ordinary user requests like processing or uploading a video, which can cause the skill to run in contexts where the user did not explicitly intend full automation. Because this skill can download content, upload to OSS, publish to external platforms, and write records, accidental triggering could lead to unintended data transfer or publication.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def run_cmd(cmd, timeout=300):
"""运行命令并返回 (returncode, stdout, stderr)"""
r = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)
return r.returncode, r.stdout, r.stderr
def ffprobe_info(video_path):
Data is uploaded to cloud storage (S3 / GCS / Azure Blob). This may be a legitimate backup or exfiltration to an external bucket. Manual review is recommended.
oss2.resumable_upload(bucket, oss_key, file_path, part_size=10 * 1024 * 1024, num_threads=4)
else:
with open(file_path, "rb") as f:
bucket.put_object(oss_key, f)
ep = OSS_ENDPOINT if OSS_ENDPOINT.startswith("http") else f"https://{OSS_ENDPOINT}"
host = urlparse(ep).netloc
The function prepares instructions for an agent to create Feishu records containing source links, OSS URLs, titles, authors, and timing metadata. This can leak sensitive user workflow data into an external SaaS system and is more dangerous because it is framed as part of the normal automation path rather than as a separately authorized export.
The pipeline sends user-provided source links, generated media, cover images, and metadata to OSS, Yixiaoer, and Feishu with only operational log messages, not meaningful consent or privacy gating. If users provide private or sensitive content, the skill can distribute that data to multiple third parties and create durable records without an explicit confirmation step.
The file presents all user-facing instructions, examples, and parameter descriptions in Chinese and does not indicate that the skill supports another language or that Chinese is a documented requirement. This can violate a language-choice policy when a skill forces a specific language without user opt-in.
Natural-language strings in the module docstring and command help are exclusively Chinese, with no indication that users may choose another language. This creates a language-policy issue when the skill is intended for general use rather than a clearly documented Chinese-only context.
The code loads cloud-storage and publishing-service secrets from environment variables to operate. While this may be operationally useful, the manifest description does not mention credential access or secret handling as part of the skill's scope, making this capability broader than the user-stated purpose alone justifies.