T01 · Skill Instruction Hijacking
- Location
SKILL.md:52- Finding
Hardcoded Feishu Recipient Redirects Skill Output to a Fixed Third Party
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is mostly coherent as an image-fetching workflow, but it can send results to a hardcoded Feishu user and processes unvalidated remote images locally.
Review this skill before installing. It should only be used if you intentionally want Xiaohongshu image retrieval, local image storage/conversion, persistent deduplication records, and Feishu delivery. Remove or parameterize the hardcoded Feishu recipient, require confirmation before sending, enforce HTTPS and host allowlists for image downloads, add file size/type checks, and avoid or sandbox ImageMagick processing of remote content.
SKILL.md:52Hardcoded Feishu Recipient Redirects Skill Output to a Fixed Third Party
daily_beauty.py:234Unvalidated MCP-Supplied URLs Enable SSRF and Unsafe Image Processing
The trigger phrases are broad, common conversational terms like '美图' and '看美女', making accidental invocation likely during ordinary chat. Because the skill performs external searching, image downloading, and state updates, unintended activation can cause undesired content retrieval and side effects without clear user intent.
The skill invokes a Python script and documents behavior that includes network access, local file reads/writes, and message sending, but it declares no explicit tool scope or permissions. This creates an over-privileged, opaque execution model where users and the platform cannot easily constrain what the skill may access or modify.
The skill description states that it downloads images and updates deduplication records, but it does not clearly warn users that local files will be modified as part of normal execution. Hidden side effects reduce informed consent and can surprise operators by persisting tracking data across runs.
The skill includes instructions to send images to a specific Feishu user via the message tool, but it provides no privacy warning or consent boundary for transmitting downloaded content to an external recipient. This can result in unauthorized sharing, misdelivery, or covert exfiltration of content and metadata outside the local environment.
The natural-language description and behavior are entirely framed in Chinese and target Chinese-platform content, with no indication that language choice is optional or that the locale restriction is a documented requirement. This can violate language/locale policy when a skill implicitly forces a specific language without user opt-in.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
def init_session(self):
"""初始化 MCP 会话"""
resp = requests.post(
MCP_URL,
headers=self.headers,
json={
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
def init_session(self):
"""初始化 MCP 会话"""
resp = requests.post(
MCP_URL,
headers=self.headers,
json={
The skill transmits search keywords and triggers retrieval of third-party platform content through a local MCP service without any consent, disclosure, or validation of the downstream service’s trustworthiness. While the endpoint is localhost, it acts as a proxy to external data sources, so this still represents outbound data flow and dependency on a local privileged broker.
def search_feeds(self, keyword):
"""搜索帖子"""
resp = requests.post(
MCP_URL,
headers=self.headers,
json={
The skill sends user/profile identifiers and tokens to a local MCP broker to fetch account profile data, creating a trust boundary with a local service that may have broader network access. If the MCP service is compromised or overly permissive, this can expose browsing targets and enable unintended data collection or abuse of platform access tokens.
def get_user_profile(self, user_id, xsec_token):
"""获取用户主页"""
resp = requests.post(
MCP_URL,
headers=self.headers,
json={
The script automatically downloads remote images into a local workspace without explicit user confirmation, content validation, or size restrictions. This creates a trust boundary issue: untrusted external content is persisted locally and may consume storage, introduce harmful files, or be processed later by other tools.
The skill invokes an external image-processing subprocess on downloaded, untrusted files without warning or isolation. This increases risk because malformed images can target vulnerabilities in the converter, and the user is not informed that local execution on remote content will occur.
The script executes an external ImageMagick convert binary on files downloaded from remote URLs. Processing attacker-controlled image content with a complex native parser expands the attack surface and can lead to command-adjacent abuse or exploitation of image parsing vulnerabilities in the converter, even though shell injection is not present here because arguments are passed as a list.
f.write(resp.content)
# 转换为 PNG
subprocess.run(
["convert", str(webp_path), str(png_path)],
capture_output=True,
check=True
The manifest describes a content-selection/push skill that searches Xiaohongshu and returns 9 images. While downloading images is an expected implementation detail, spawning a local system executable (convert) introduces host-level execution capability that is not implied by the stated purpose.
The manifest describes returning 9 full-body images for one blogger when triggered, but this data file holds hundreds of image URLs accumulated over time. That indicates the skill persists a historical corpus of fetched images, which is broader behavior than the user-facing description suggests.
For a skill whose stated purpose is to search and return one blogger's 9 images on request, keeping a long-lived local index of many remote image URLs adds data-retention capability beyond simple request-time retrieval. This persistence can enable tracking, replay, or redistribution of a much larger set of images than the manifest describes.
All image URLs use plain HTTP, which allows network attackers to intercept or modify the image responses in transit. In a skill that renders remote media to users, this creates a realistic content-tampering and tracking surface, and could expose users to malicious or inappropriate substituted content if the client fetches these links directly.
This markdown file presents all skill content in a single language and does not indicate that the user can choose another language or that the skill is intentionally limited to a Chinese-language context. Under the policy criteria, forcing a specific language without user opt-in is a natural-language policy concern.
The entire whitelist document is written in Chinese and provides no indication that language selection is optional or that the skill is intentionally limited to Chinese-speaking users. Under the stated policy, forcing a specific language without user opt-in can be a natural-language policy violation.