T09 · Insecure Skill Coding Practices
- Location
SKILL.md:52- Finding
Shell Command Injection Through Unquoted Dynamic Arguments
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This shopping skill is purpose-aligned but needs Review because it can place real or scheduled orders while using broad triggers, plaintext credentials/address data, unsafe command patterns, and weak transaction safeguards.
Install only after reviewing the API endpoint and replacing packaged credentials/address data. Require explicit invocation and final confirmation before any order, disable or tightly control scheduled purchases and auto-push, avoid full address disclosure in chat, and fix the command/path/input-validation issues before using it with real payment authority.
SKILL.md:52Shell Command Injection Through Unquoted Dynamic Arguments
scripts/seap-cli.js:75Arbitrary Cache File Access Through Session ID Path Traversal
seap.config.json:2Plaintext Credential-Like Value and Delivery Address Included in Distributed Configuration
scripts/seap-cli.js:173Insufficient Order Validation and Hardcoded Payment Identity Fields
The README states the skill auto-triggers whenever a user message contains the single Chinese character "买". For a skill that can search, purchase immediately, and schedule future purchases, such a broad trigger creates a serious risk of accidental activation from ordinary conversation, leading to unintended commerce flows or purchases.
The README describes stored payment credentials (secretToken) and automated scheduled purchasing, but does not prominently warn users about irreversible financial actions, credential sensitivity, or the risk of unattended purchases. In the context of a shopping skill, this materially increases the chance of unauthorized or accidental financial transactions.
The manifest describes a shopping skill, but the actual behavior goes further by handling local persistence, authentication credentials, network/API communication, and personal address/order data. That mismatch can mislead reviewers and users about the sensitivity of the skill, causing them to approve or invoke a workflow with greater privacy and financial risk than advertised.
Auto-triggering on any message containing the character “买” is overly broad and can activate the purchase workflow from casual discussion, quoted text, or unrelated messages. In a skill capable of ordering goods and scheduling unattended purchases, accidental activation materially raises the risk of unwanted data exposure and unintended transactions.
Referenced artifact was not completely inspected
依赖 `scripts/seap-cli.js` 执行相关命令。技能可在普通环境中运行,无需acpx运行时。
The skill supports scheduled automatic purchasing but does not present a strong, upfront warning that the transaction may be executed later without the user being present and may be difficult or impossible to reverse. In a financial workflow, unattended execution significantly increases the risk of accidental or stale-intent purchases.
The skill description is presented as a Chinese-only shopping skill and all example interactions are in Chinese, with no indication that users may choose another language or that the locale limitation is intentional and justified. This can violate language/locale policy when a skill imposes a specific language without opt-in.
The skill invokes a CLI that performs network-backed search and purchase operations and reads/writes local files, but it declares no explicit tool scope or permission boundaries. In a purchasing skill handling auth material, addresses, and order execution, missing capability declarations weakens reviewability and increases the chance of unintended or excessive tool access.
The manifest trigger phrase is vague and does not clearly constrain when this high-impact purchasing skill should run. Ambiguous activation criteria increase the chance that the skill is invoked outside the user's intended context, which is especially risky when financial actions and personal data are involved.
The configuration supports automatic result pushing to a messaging channel, but the skill text does not clearly warn that purchase results and address-related information may be sent automatically. Without explicit notice and opt-in, sensitive order metadata may be disclosed through channels the user did not expect.
The instructions require echoing the user's full saved shipping address back into chat. Displaying full address and phone details in chat increases the chance of exposure to logs, observers, shared-session viewers, or downstream notification systems.
The workflow instructs the agent to update seap.config.json with user-supplied address information, but it does not prominently warn the user that their address will be written to a local configuration file for future reuse. Silent persistence of personal data increases privacy risk and can surprise users who intended a one-time purchase only.
The scheduled-task design instructs the agent to push purchase results, including address and order information, into another session. Cross-session delivery of sensitive order data expands the exposure surface and can leak personal information if session targeting is wrong, reused, or insufficiently verified.
The scheduled-purchase flow gives contradictory instructions about whether results are automatically shown in the current session or must be pushed with sessions_send to the original session. In security-sensitive purchase automation, this inconsistency can cause result leakage to the wrong session, missed notifications, or duplicate delivery of order and address data.
The aipay command assembles purchase data including shipping address, SKU, merchant ID, authCode, and payment-related fields, then sends it to a remote host without any interactive confirmation, consent prompt, or clear disclosure at execution time. In a shopping skill, this directly enables unintended purchases and disclosure of sensitive personal/order data if the command is triggered unexpectedly or by a higher-level agent without explicit user approval.
The search command sends the user's search intent together with an authCode to a remote server, but the CLI provides no visible notice, consent, or data-use disclosure before transmission. While search is less sensitive than order placement, the authCode is a credential-like secret and the search intent may reveal private interests or behavior, making silent transmission risky.