T01 · Skill Instruction Hijacking
- Location
scripts/main.py:156- Finding
Persistent Commercial Output Injection and Payment Redirection
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This paid prompt-search skill has real review concerns because it sends user queries and IDs to a hardcoded external service, redirects users to a payment page, and requests broad execution authority.
Review this skill before installing. It is not showing destructive or backdoor behavior, but users should understand that searches and user identifiers go to a third-party endpoint, access is tied to an external payment flow, and the manifest asks for exec despite no clear need in the visible code. Install only if you trust the publisher, endpoint, and payment destination, and avoid entering confidential business ideas unless the data handling terms are clear.
scripts/main.py:156Persistent Commercial Output Injection and Payment Redirection
scripts/main.py:20Search Queries and User Identifiers Disclosed to an External Gateway
scripts/main.py:220Untrusted Remote Content Rendered Without Markdown Sanitization
skill.yml:46Unpinned Third-Party Python Dependency
The skill is presented as a prompt-retrieval assistant, but its documented workflow also generates user IDs, gates content on registration/payment state, and directs users to a third-party payment flow. This mismatch is dangerous because it hides material behavior from both users and platform reviewers, increasing the risk of deceptive monetization, unauthorized data handling, and misuse of the skill as a payment funnel.
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
elif isinstance(item, str):
if len(item) > 200 and 'Role' in item:
prompts.append(item)
return prompts
# data 是字符串
if not data or not isinstance(data, str):
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
elif isinstance(item, str):
if len(item) > 200 and 'Role' in item:
prompts.append(item)
return prompts
# data 是字符串
if not data or not isinstance(data, str):
The trigger phrases include extremely broad terms such as 想要 and 获取, which are common in normal conversation and can cause accidental or adversarial activation outside the intended e-commerce prompt-retrieval context. Over-broad activation increases the chance that the skill is invoked on unrelated user content, potentially exposing privileged tools or returning unintended prompt material.
The skill describes calling an external LAF API but does not declare any tool scope or allowed network permissions. This creates a transparency and governance gap: the host or reviewer cannot clearly constrain or audit outbound access, and users are not informed that their queries may be transmitted externally.
The description, triggers, usage instructions, and output templates are entirely in Chinese and present the skill's interaction model in a single language with no opt-in or alternative locale support. Under the policy, forcing a specific language without user choice is a natural-language policy violation unless the locale restriction is explicitly justified.
Several trigger patterns are broad natural-language phrases and regex-like forms such as '想要.*的提示词' and '获取.*的架构', which can match many ordinary requests outside the intended narrow skill scope. Overbroad activation can cause unintended invocation, sending unrelated user queries into external retrieval or payment-gated flows without clear intent.
The workflow states that the skill will automatically generate a user ID and perform external API/payment processing, but it does not provide a clear user-facing disclosure about what data is collected, where it is sent, or how it is used. In context, this is more dangerous because the skill combines identity generation, search telemetry, and third-party payment redirection, which can expose users to privacy loss and social-engineering-style monetization without informed consent.
A hardcoded payment URL is embedded directly in a prompt-search skill, which is not necessary for simple retrieval functionality. In this context, the skill can steer users to an external payment flow from within the agent, increasing the risk of deceptive monetization or abuse if the endpoint changes or is not independently verified.
The code transmits the user's search query and user ID to an external service without any in-file disclosure or consent mechanism. Because queries may contain sensitive commercial ideas or personal identifiers, this creates a privacy and data-governance risk, especially when the service is third-party hosted.
This outbound request sends user-supplied query data and a user identifier to an external host. In a skill that claims to retrieve prompts, such transmission is contextually sensitive because users are not clearly informed that their input and account linkage are being shared off-platform.
"""单次搜索调用"""
payload = {"action": "search", "query": query, "user_id": user_id}
try:
response = requests.post(LAF_URL, json=payload, timeout=10)
return response.json()
except Exception:
return None
The skill goes beyond its stated purpose of searching and returning e-commerce prompts by implementing registration state handling, paywall enforcement, payment-link distribution, and activation messaging. This is risky because it can monetize or gate access through the agent interface without clear platform-level disclosure or trust guarantees, creating phishing, policy, and user-manipulation concerns.
All user-facing status and instruction messages in this file are presented in Chinese, including registration, payment, and result text, with no apparent option for the user to choose another language. That constitutes a language/locale constraint in natural-language behavior without documented opt-in or justification.
The payment-status check sends the same query and user ID to the external endpoint again, even though only account status appears necessary. This expands unnecessary data exposure and couples access control with content queries, increasing privacy risk and data collection beyond what is needed.
This second outbound request duplicates transmission of query and user ID to the same external service for status checking. The duplicate transfer is unnecessary for least-privilege data handling and broadens third-party visibility into user behavior and billing-related metadata.
# ==================== 步骤3:检查支付状态 ====================
check_payload = {"action": "search", "query": query, "user_id": user_id}
try:
check_response = requests.post(LAF_URL, json=check_payload, timeout=10)
check_data = check_response.json()
check_status = check_data.get("status") or check_data.get("ec")
except Exception:
The visible name and description are entirely in Chinese, with no indication that users may choose another language or that the skill is intentionally restricted to a Chinese-speaking audience for a documented reason. This can constitute a language/locale policy issue because the skill presentation implicitly enforces one language without opt-in or justification.
The invocation configuration defines positive triggers and regex patterns but provides no scope boundaries, exclusions, or examples of when the skill must not run. In combination with broad triggers, this makes misrouting more likely and can route unrelated user input into a skill that may possess elevated capabilities, which is especially concerning here because the manifest also requests exec.
The manifest requests the exec capability even though the skill is described as a prompt/search retrieval assistant and does not document any legitimate need for shell execution. Unnecessary command-execution privileges materially expand the attack surface: if later skill logic interpolates user-controlled input into shell commands, this can enable arbitrary command execution, data access, or lateral abuse of the host environment.