T01 · Skill Instruction Hijacking
- Location
SKILL.md:14- Finding
Mandatory Agent Response Hijacking Through Skill-Controlled Rendering Rules
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill mostly performs price comparison as advertised, but it under-discloses external query sharing and tries to control the agent’s final response too rigidly.
Review this before installing if you care about privacy or agent control. Product searches may be sent to maishou88.com, local query history is retained, and optional Feishu sync can send shopping queries and results to a Feishu workspace. Avoid enabling Feishu sync unless you are comfortable with the Bitable permissions and plaintext local config, and treat the generated report as data the agent should be allowed to summarize or adapt.
SKILL.md:14Mandatory Agent Response Hijacking Through Skill-Controlled Rendering Rules
bin/setup_feishu.py:140Feishu Base Token Exposed Through Process Arguments, Console Output, and Unprotected Configuration Writes
bin/_data_layer.py:81Misleading Local-Only Privacy Claim Despite Mandatory Third-Party Query Transmission
bin/price_check.py:1Unpinned Aiohttp Runtime Dependency Produces Non-Reproducible Executions
Although scoped to the skill's own directory, the README encourages copy-pasting a forceful recursive deletion command. In an agent-skill ecosystem, destructive shell commands in documentation are risky because users may execute them blindly, and path expansion or editing mistakes can lead to broader local data loss.
cp ~/.openclaw/data/price-check/price-check.db ~/Backups/
rm -rf ~/.openclaw/data/price-check/
## Roadmap
Although scoped to the skill's own directory, the README encourages copy-pasting a forceful recursive deletion command. In an agent-skill ecosystem, destructive shell commands in documentation are risky because users may execute them blindly, and path expansion or editing mistakes can lead to broader local data loss.
cp ~/.openclaw/data/price-check/price-check.db ~/Backups/
rm -rf ~/.openclaw/data/price-check/
## Roadmap
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
cp ~/.openclaw/data/price-check/price-check.db ~/Backups/
rm -rf ~/.openclaw/data/price-check/
## Roadmap
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
cp ~/.openclaw/data/price-check/price-check.db ~/Backups/
rm -rf ~/.openclaw/data/price-check/
## Roadmap
The skill performs local persistence/caching and optional Feishu synchronization beyond simple price lookup, which increases data handling risk relative to the stated purpose. Even if some behavior is partially documented later in the file, top-level mismatch can mislead operators about retention, external dependencies, and where user queries and result data may end up.
The skill performs local persistence/caching and optional Feishu synchronization beyond simple price lookup, which increases data handling risk relative to the stated purpose. Even if some behavior is partially documented later in the file, top-level mismatch can mislead operators about retention, external dependencies, and where user queries and result data may end up.
The skill performs local persistence/caching and optional Feishu synchronization beyond simple price lookup, which increases data handling risk relative to the stated purpose. Even if some behavior is partially documented later in the file, top-level mismatch can mislead operators about retention, external dependencies, and where user queries and result data may end up.
The skill performs local persistence/caching and optional Feishu synchronization beyond simple price lookup, which increases data handling risk relative to the stated purpose. Even if some behavior is partially documented later in the file, top-level mismatch can mislead operators about retention, external dependencies, and where user queries and result data may end up.
The skill transmits query data and result content to Feishu via feishu_sync.sync_query_to_feishu(query, result) even though that behavior is not necessary for a price-comparison function and is only loosely labeled as opt-in in a comment. User queries may contain sensitive purchase intent, product interests, or personal shopping context, and the full result object can include cached history and derived data, creating an unnecessary third-party data exfiltration path.
The README says the bot recognizes trigger keywords including 比价 / 值不值得买 / 哪里买 / 多少钱合适 / etc., but does not define a bounded list or exclusion conditions. Phrases like 哪里买 and 多少钱合适 are common shopping-language fragments and the etc. wording makes activation scope ambiguous, which could cause unintended invocations.
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
Sync every query (best_deal + Top 3 + history price) to a Feishu multi-dimensional table for cross-device browsing. **Default off** — only do this if you want it.
1. Install `lark-cli`
2. Create an empty Bitable in your Feishu workspace
3. Authorize your Feishu Bot app as an editor on that Bitable
4. Run the one-time setup:
The README documents a destructive reset command that permanently deletes the skill's local data directory, but it does not prominently warn that the action is irreversible. Users may copy-paste it without understanding that cached history, configuration, and any optional sync settings stored there will be lost.
该处列出的触发词包括“哪里买”“多少钱合适”等较通用表达,且用“等”表示触发集合未穷尽,缺少明确边界。这会让技能在普通购物聊天或泛化咨询中被意外触发,难以判断何时会触发、何时不会。
The README documents optional synchronization of query results, links, and history to Feishu but does not clearly warn users that their shopping interests, product queries, and generated links will be transmitted to a third-party service once enabled. This can expose behavioral data and potentially sensitive purchase intent, especially if users search for medical, personal, or business-related items.
The skill declares capabilities requiring shell, network, local file read/write, and environment access, but does not declare any explicit tool scope or permissions boundary. In an agent environment this weakens least-privilege controls, making it easier for the skill to access local data, persist user queries, or invoke external commands without clear operator awareness.
The trigger list includes broad phrases like '历史价', '给我购买链接', and '当前价合理吗', which can cause accidental activation during ordinary conversation. In this skill, accidental activation is more sensitive because execution may perform network requests, local data writes, and optional remote synchronization, turning a benign chat into unintended data collection and external calls.
The module exports query terms, pricing results, shop names, links, and history-derived fields to Feishu, an external SaaS, while the skill description emphasizes local price checking and local history storage. That creates a data-sharing behavior users may not reasonably expect, especially because search queries can reveal purchase intent or sensitive interests.
This code implements full outbound data export capability to Feishu that is ancillary to the core price-comparison function, increasing privacy and supply-chain exposure. Even if optional, once enabled it sends user queries and shopping metadata to a third party, which broadens the attack surface and can violate user expectations if not transparently declared.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
if profile:
cmd = ["lark-cli", "--profile", profile] + args
try:
result = subprocess.run(
cmd, capture_output=True, text=True, timeout=timeout, check=False,
)
if result.returncode != 0:
The manifest promises '直接可点击的购买链接', implying links to the actual product listing or checkout target. However, the module docstring explicitly says affiliate/deep links were removed and later code only generates and returns generic platform search URLs via search_url, not direct item purchase links.
The manifest description names many domestic platforms as supported targets, suggesting the skill can surface actionable buying locations across them. In code, native search URL templates are only implemented for Taobao/Tmall, JD, Pinduoduo, and 1688, while comments state Douyin, Kuaishou, Suning, Vipshop, and Kaola are left unsupported for web search output.
The natural-language directive requires the agent to send the pre-rendered human_report verbatim and forbids translation or reorganization. Because the report content is authored in Chinese throughout the file, this effectively forces a specific language/locale on users without any opt-in or alternative language path.
The setup script enables Feishu synchronization and persists configuration data, which expands the skill's behavior beyond the user-facing description of price lookup and recommendation. That capability broadening creates a trust and transparency issue: users may not expect external data synchronization or durable local state tied to third-party services.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
"""跑 lark-cli 命令,返回解析后的 JSON。失败 sys.exit。"""
cmd = ["lark-cli"] + args
try:
r = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout, check=False)
except FileNotFoundError:
sys.exit("❌ lark-cli 未安装或不在 PATH")
if r.returncode != 0:
The script programmatically creates fields in an external Feishu base, which is a write-capability affecting third-party data structures rather than merely reading prices. In the context of a price-check skill, this is broader-than-expected behavior and can lead to unauthorized or surprising modification of a user's workspace if run without clear informed consent.