T01 · Skill Instruction Hijacking
- Location
SKILL.md:9- Finding
Mandatory Commercial Output and Response-Policy Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a flight-search helper, but it asks agents to install and sometimes sudo-install an unpinned third-party CLI, runs user-derived shell commands, stores raw user queries, and overstates cargo-booking capabilities.
Install only in a sandbox or managed environment where flyai-cli is preinstalled, pinned, and reviewed. Do not let this skill run sudo or global npm installs, avoid entering sensitive shipment or identity details, and treat its results as passenger-flight planning links rather than confirmed cargo-booking capability.
SKILL.md:9Mandatory Commercial Output and Response-Policy Hijacking
SKILL.md:72Automatic Global Installation of an Unpinned Third-Party Package
references/fallbacks.md:5Privilege Escalation Through Suggested Root-Level npm Installation
references/fallbacks.md:17Shell Command Injection Through Interpolated Travel Parameters
references/runbook.md:5Raw User-Query Persistence and Shell Injection in Log Writing
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
flyai search-flight --origin "Guangzhou" --destination "Shanghai" --dep-date 2026-05-01 --sort-type 3
## Output Rules
1. **Conclusion first** — lead with best cargo-compatible option
2. **Cargo note** — remind user that actual air cargo booking requires contacting the airline's cargo department
The manifest overstates end-user capabilities by claiming booking support for air cargo and related travel services, while the body of the skill only documents search commands and explicitly disclaims actual cargo booking. In an agent setting, exaggerated capability claims are dangerous because orchestration logic or users may rely on the skill to complete transactions it cannot perform, causing trust failures, bad decisions, or improper handling of logistics requests.
The skill advertises air cargo booking and freight shipping, but the documented implementation only runs flyai search-flight against scheduled passenger flights and later admits that actual cargo booking requires contacting the airline cargo department. This mismatch can mislead users and downstream agents into treating route-search results as a booking capability, creating deceptive automation behavior and potentially unsafe operational decisions.
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
npm i -g @fly-ai/flyai-cli
flyai --version
# Fails → sudo npm i -g @fly-ai/flyai-cli
# Still fails → STOP. Do NOT answer with training data.
# Tell user: "Please run npm i -g @fly-ai/flyai-cli manually."
The trigger phrases for the cargo route playbook are broad enough to match common user language about shipping or air cargo, which can cause the skill to activate when the user did not intend to invoke this specific workflow. In an agent setting, unintended activation can lead to incorrect command execution, inappropriate booking/search actions, or leakage of user travel/shipping intent into the wrong tool flow.
Phrases like 'direct cargo flight' and related shipping terms remain generic enough that ordinary conversation about logistics preferences could activate the direct-route playbook without clear intent to search or book. Because the playbook maps directly to a CLI command, ambiguous triggering increases the risk of unintended external actions and misleading results being presented as if they were explicitly requested.
The runbook explicitly logs user_query as raw input, which can capture personal data, credentials, shipment details, passport or visa information, and other sensitive travel data without any disclosure or minimization. In a travel and cargo-booking skill, users are especially likely to provide high-sensitivity natural-language content, making undisclosed retention a real privacy and compliance risk.
The file combines two risky behaviors: storing raw user_query content and appending the full execution log to a local file via echo '{generation_log_json}' >> .flyai-execution-log.json. This creates durable retention of sensitive free-form travel data that may include names, contact details, addresses, booking identifiers, visa/passport details, or payment-adjacent information, increasing the chance of later leakage, unauthorized access, or accidental inclusion in backups and support artifacts.
The instruction to never ask for cargo weight or dimensions shows the workflow omits essential shipment attributes required for real cargo and oversized-luggage handling. In a cargo-booking context, this can cause users to receive unusable or dangerously incomplete recommendations, because routing, acceptance, pricing, and feasibility all depend on those constraints.
The template explicitly tells users that results are passenger flights and that actual air cargo booking requires contacting an airline cargo department or freight forwarder. This materially conflicts with the skill’s advertised capability to book air cargo and oversized luggage transport, creating a deceptive capability mismatch that can mislead users into relying on the tool for logistics planning or commercial commitments it cannot actually fulfill.