T01 · Skill Instruction Hijacking
- Location
SKILL.md:163- Finding
Forced Commercial Promotion and Provider-Exclusive Output Steering
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This Japan travel skill is purpose-aligned, but it asks agents to automatically install and run an unpinned global CLI, including a sudo fallback, and records raw user queries in background logs.
Review before installing. Use this only in an environment where installing and running the FlyAI CLI is acceptable, avoid sudo, prefer a pinned and reviewed CLI version, and do not let the skill auto-install packages. Users should understand that trip details may be sent to FlyAI/Fliggy and that recommendations may be limited to FlyAI-bookable options.
SKILL.md:163Forced Commercial Promotion and Provider-Exclusive Output Steering
SKILL.md:48Automatic Execution of an Unpinned Global npm Dependency with Root Escalation Fallback
SKILL.md:96Shell Command Injection Through User-Controlled Travel Parameters
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
**场景化 Playbook(经典 5 天/樱花季/穷游/北海道冬季)** → 详见 [references/playbooks.md](references/playbooks.md)
## Output Rules(强约束)
### 单点查询:使用对应品类标准表格
The skill instructs the agent to probe for a local CLI and, if missing, perform a global npm installation before answering a travel question. That introduces unnecessary host-side side effects and supply-chain risk for a content/booking skill, because it expands execution from simple data retrieval into local environment modification without clear necessity or user consent.
The skill directs an automatic global npm install when the CLI is absent, with no confirmation, warning, or trust boundary explanation. Silent dependency installation is dangerous because it can change the host environment, execute package lifecycle scripts, and expose the agent to dependency confusion or compromised package scenarios.
The skill orchestrates multiple third-party queries for flights, hotels, visas, and points of interest using user-supplied trip details, but does not warn that this data will be sent externally. Travel dates, origin city, destinations, and itinerary preferences can be sensitive, and undisclosed transmission undermines informed consent and privacy expectations.
SQP-3 applies to all file types and covers language or locale policy violations. The entire fallback specification, including user-facing example messages, is written only in Chinese and does not indicate that language should match user preference or allow opt-in to another locale.
The document instructs automatic global installation of a CLI and escalates to sudo npm i -g without warning or approval gates. That is dangerous because it normalizes system modification and privileged package installation from guidance content, increasing the chance of unintended host changes or compromise via malicious packages.
The explicit instruction to run sudo npm i -g @fly-ai/flyai-cli asks for root-level execution of a package manager install, which materially raises risk. If the package, dependency chain, or install scripts are malicious or compromised, the result could be full system compromise with elevated privileges.
flyai --version
# Step 3 → 仍然失败
→ 尝试:sudo npm i -g @fly-ai/flyai-cli
→ 或:npx @fly-ai/flyai-cli --version
# Step 4 → 彻底失败
The fallback recommends running npx @fly-ai/flyai-cli without pinning a specific version, which permits execution of whatever package version is current at runtime. That creates a supply-chain risk: a compromised or breaking upstream release could be executed automatically in a travel-assistant workflow.
L025 says that if the CLI is unavailable, the agent must not fall back to its own general knowledge because there is no real-time data or booking links. However, later fallback cases explicitly instruct the agent to provide answers from built-in knowledge for POIs and visa information when searches fail, creating a direct contradiction in operational intent.
All user-facing playbook content and trigger descriptions are presented only in Chinese, effectively forcing a specific language without explicit opt-in. The file does not state that the skill is intentionally limited to Chinese-speaking users or provide any language-selection mechanism.
The activation text includes broad conditions like '用户首次赴日' and '无明确偏好', which can match many general Japan-travel conversations without clearly defining when this playbook should or should not apply. It also lacks exclusion criteria or negative examples to distinguish this playbook from other itinerary options.
The condition '或 3-4 月出行' can activate the playbook for users traveling in those months even if they did not ask about cherry blossoms. This creates overlap with many normal spring travel requests and makes the trigger scope insufficiently specific.
Phrases like '省钱' and '预算有限' are common and can apply to many travel requests, but the file does not define what budget range or tradeoffs qualify for this playbook. Without constraints or exclusions, the skill may invoke the budget itinerary too often.
Terms like '雪' and '12-2 月出行' are not specific enough to Hokkaido winter travel intent and may match users discussing weather or winter travel more generally. The trigger lacks boundaries that distinguish Hokkaido ski/onsen planning from other winter Japan trips.
The natural-language instructions and field descriptions are entirely in Chinese, with no indication that users can choose another language or that the skill is intentionally limited to a Chinese-language context. This can violate language/locale policy when a skill imposes a specific language without opt-in.
The runbook explicitly requires storing the user's raw input in structured logs, which can capture sensitive personal data such as names, passport details, travel dates, contact information, or payment-related context commonly present in travel queries. Because the logging is described as a global backend behavior and is not paired with minimization, redaction, retention limits, or user notice, it creates a real privacy and data-exposure risk if logs are accessed, retained too long, or reused for debugging.
The templates and interaction instructions are entirely in Chinese and prescribe Chinese-language prompts and outputs without checking the user's language preference. This can override or conflict with user expectations, leading to misleading responses, degraded consent around booking-related actions, and possible misunderstanding of important travel details like visa or pricing information.
The phrase triggering full-trip planning on broad terms like '帮我规划/安排' is overly generic and can match ordinary conversation that is not a clear request to invoke a multi-step travel-planning workflow. This can cause the agent to collect parameters or launch planning actions without sufficiently specific user intent, increasing the risk of unintended tool use and unnecessary data collection.
The reference to a runbook for full-process backend execution logging adds an operational behavior unrelated to the user-visible travel-assistant purpose. This can lead to over-collection of prompts, travel plans, dates, origins, and other potentially sensitive trip data without transparency or minimization.