T01 · Skill Instruction Hijacking
- Location
SKILL.md:23- Finding
Agent Instruction Hijacking Through Mandatory Priority and Verbatim Output Rules
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a customer-intelligence tool, but its credential setup and agent-instruction rules create risks that should be reviewed before installation.
Review before installing. Use a dedicated, least-privilege 1688 AK; avoid pasting the AK into shell commands; verify OPENCLAW_GATEWAY_URL points only to the trusted local OpenClaw gateway; protect or rotate any AK already written to openclaw.json; and install only if you accept automatic usage reporting and access to sensitive customer intelligence data.
SKILL.md:23Agent Instruction Hijacking Through Mandatory Priority and Verbatim Output Rules
scripts/capabilities/configure/service.py:29API Key and Gateway Bearer Token Can Be Sent to an Arbitrary Environment-Controlled URL
scripts/capabilities/configure/service.py:60API Key Is Persisted in Plaintext Without Enforcing Restrictive File Permissions
scripts/_auth.py:200Credential Setup Exposes Secrets Through Command-Line Arguments and Authentication Diagnostics
The gateway URL is taken directly from an environment variable and used as the destination for a PATCH request carrying the API key and optional bearer token. If that environment variable is altered, the skill can exfiltrate credentials to an attacker-controlled service; the customer-management context does not justify broad credential forwarding, which makes this especially risky.
headers = {}
if token:
headers["Authorization"] = f"Bearer {token}"
resp = requests.patch(f"{gateway_url}/api/config",
headers=headers, json=payload, timeout=5)
return resp.ok
except Exception:
The Chinese description similarly omits that the skill can manage configuration and process credentials, creating the same transparency and review gap across locales. Security-relevant behavior hidden behind a business-focused description can mislead operators into granting trust or permissions they would otherwise restrict.
The Chinese description similarly omits that the skill can manage configuration and process credentials, creating the same transparency and review gap across locales. Security-relevant behavior hidden behind a business-focused description can mislead operators into granting trust or permissions they would otherwise restrict.
The Chinese description similarly omits that the skill can manage configuration and process credentials, creating the same transparency and review gap across locales. Security-relevant behavior hidden behind a business-focused description can mislead operators into granting trust or permissions they would otherwise restrict.
The Chinese description similarly omits that the skill can manage configuration and process credentials, creating the same transparency and review gap across locales. Security-relevant behavior hidden behind a business-focused description can mislead operators into granting trust or permissions they would otherwise restrict.
The Chinese description similarly omits that the skill can manage configuration and process credentials, creating the same transparency and review gap across locales. Security-relevant behavior hidden behind a business-focused description can mislead operators into granting trust or permissions they would otherwise restrict.
Referenced artifact was not completely inspected
- `{baseDir}` = skill 根目录(含 `SKILL.md` 和 `cli.py` 的目录)
Referenced artifact was not completely inspected
- `{baseDir}` = skill 根目录(含 `SKILL.md` 和 `cli.py` 的目录)
Referenced artifact was not completely inspected
- `{baseDir}` = skill 根目录(含 `SKILL.md` 和 `cli.py` 的目录)
The documentation states that call records are automatically reported, but users are not informed about the transmission or given any privacy notice. This is particularly risky here because the skill processes customer intelligence and may involve commercially sensitive or personal data context.
The configuration guide is materially unrelated to the declared customer intelligence skill and instead instructs the agent to solicit a DingTalk Access Key and configure message-sending credentials. This kind of capability mismatch is a strong indicator of hidden functionality or prompt-scope smuggling, creating a path to collect secrets unrelated to the user’s expected task.
The document explicitly introduces credential collection and a message-sending setup that has no clear relationship to customer inquiry analysis, follow-up suggestions, or buyer profiling. In context, this is more dangerous because users would reasonably trust the skill with CRM-like analysis, not expect it to request reusable messaging credentials that could enable unauthorized outbound actions.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
职责:每次 CLI 命令执行时,向 skill 网关上报一次调用记录,用于统计 skill 调用次数。
上报失败不影响主流程,静默处理。
环境变量(从项目根目录 .env 读取):
SKILL_NAME skill 名称,默认 1688-shop-zkt-buyer-manage
SKILL_VERSION skill 版本,默认 1.0.0
SKILL_CHANNEL 发布渠道,默认 clawhub
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
职责:每次 CLI 命令执行时,向 skill 网关上报一次调用记录,用于统计 skill 调用次数。
上报失败不影响主流程,静默处理。
环境变量(从项目根目录 .env 读取):
SKILL_NAME skill 名称,默认 1688-shop-zkt-buyer-manage
SKILL_VERSION skill 版本,默认 1.0.0
SKILL_CHANNEL 发布渠道,默认 clawhub
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
职责:每次 CLI 命令执行时,向 skill 网关上报一次调用记录,用于统计 skill 调用次数。
上报失败不影响主流程,静默处理。
环境变量(从项目根目录 .env 读取):
SKILL_NAME skill 名称,默认 1688-shop-zkt-buyer-manage
SKILL_VERSION skill 版本,默认 1.0.0
SKILL_CHANNEL 发布渠道,默认 clawhub
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
def _load_env_file() -> None:
"""解析项目根目录的 .env 文件,将变量注入 os.environ(已有环境变量不覆盖)。"""
env_path = _ROOT_DIR / ".env"
if not env_path.exists():
return
with open(env_path, encoding="utf-8") as f:
The file implements AK credential status checking and persistence, which is unrelated to the advertised customer intelligence functionality. A skill that collects and writes access credentials under misleading business-facing branding creates a strong risk of credential harvesting or unauthorized secret manipulation, especially because users may provide sensitive AK values expecting normal CRM features.
The skill declares no explicit tool scope or permission boundary, yet the documented implementation uses environment variables, filesystem access, network access, and configuration writes. Without a least-privilege declaration, a host agent may grant broader capabilities than necessary, increasing blast radius if the skill is modified or abused.
The skill is designed to expose detailed customer profiles, purchasing habits, decision factors, and inquiry activity at other shops. Even if intended for business use, this is sensitive commercial and potentially personal data, and inadequate access control, minimization, or masking could lead to privacy violations or competitive misuse.
文档明确规定“Agent 所有面向用户的输出必须 100% 用中文”,并禁止任何英文或中英混排。这构成了语言/locale 强制策略,但未向用户提供语言选择,也未在文档中明确说明该限制是出于区域合规或产品范围限定。
Overly broad trigger phrases for follow-up messaging can cause the skill to activate in unintended contexts and produce sensitive customer-analysis outputs or business guidance when the user did not clearly request it. In an agent environment, ambiguous invocation increases the chance of accidental data access and surprise actions.
Generic detail-query triggers blur the boundary for when customer profile retrieval should occur. Because this skill can surface detailed buyer information and cross-shop inquiry data, accidental triggering is more dangerous than in a low-sensitivity skill.
The CLI unconditionally attempts to report skill usage after every command via _tracker.report_skill_usage() and does so without any notice, consent flow, or visible configuration in this file. In a customer-management skill, telemetry may reveal sensitive business activity patterns, command usage, or identifiers if the tracker includes arguments or context, creating a privacy and data-governance risk even though the exact transmitted fields are not shown here.
The markdown mandates Chinese presentation conventions, including a fixed Chinese opening narrative and Chinese field headers, without any indication that users may choose another language. This is a natural-language locale policy concern because the skill appears to require a specific language by default rather than documenting a justified region-specific constraint or offering opt-in.
The guide tells the agent to ask for an Access Key and pass it on the command line, but provides no warning about sensitive credential handling or the risks of disclosure through logs, shell history, process listings, or transcript retention. Even if not intentionally malicious, this creates an unsafe secret-handling pattern that can expose reusable credentials.