T03 · Remote Payload Retrieval and Execution
- Location
SKILL.md:58- Finding
Unpinned Remote Source Is Retrieved and Installed as a Trusted Proxy
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is openly a fast command router, but it installs unreviewed remote code and can turn ordinary chat phrases into immediate shell or service-management commands without clear safety gates.
Review carefully before installing. Only use this behind strong local access controls, pin and verify the implementation source, avoid raw shell templates, require explicit command prefixes and confirmations for state-changing routes, and run the proxy as a dedicated low-privilege user.
SKILL.md:58Unpinned Remote Source Is Retrieved and Installed as a Trusted Proxy
routes.example.yaml:50User-Controlled Message Data Is Inserted into Shell Command Templates
The description emphasizes speed and perfect execution while omitting any warning that matched inputs can directly execute scripts and operational commands without LLM mediation. This can mislead users into deploying the middleware in front of an LLM without appreciating that it creates an immediate command-execution path for matched phrases.
The examples show direct shell execution and service-management commands like 'bash /opt/scripts/deploy.sh {{service}} staging' and 'systemctl restart {{service}}' without adjacent warnings about authorization, validation, shell-safety, or operational blast radius. In context, this is especially dangerous because the product's core feature is bypassing LLM inference and executing matched commands immediately, so readers may copy unsafe patterns into production.
The skill is explicitly designed to intercept natural-language commands like deploy, restart, print, and check logs using regex and bypass LLM validation entirely. Because these are common conversational phrases and the documentation does not describe strong namespace, confirmation, or context constraints, normal user text could be misclassified as an executable command and trigger operational actions unexpectedly.
The use-case examples include very generic phrases such as 'restart nginx', 'turn on lights', and 'refund order ORD-789', which are plausible in ordinary conversation and could overlap with benign chat. In a middleware that automatically executes scripts on regex match, this broad trigger surface materially increases the chance of unintended execution of destructive or sensitive actions.
This route maps a natural-language deploy command directly to a production-capable shell action with no confirmation, approval gate, or visible safety interlock. In a zero-latency router that bypasses LLM reasoning, accidental matches or unauthorized use can trigger immediate high-impact operational changes, including production deployments.
The restart-service route activates on a very short pattern like 'restart ' (and its Chinese equivalent) without additional context, exclusions, or channel constraints. In a manifest-style routing file, this can cause unintended invocation because the trigger is a common imperative phrase rather than a narrowly scoped command form.
The configuration performs an immediate 'systemctl restart' from a simple chat-style command without warning or confirmation. Given the skill's design goal of intercepting known commands and executing directly, this increases the chance of disruptive service restarts from mistaken input or insufficiently authorized users.
The generate-report route allows broad verbs like 'generate', 'create', and especially 'run' followed by 'report ', which can be interpreted in multiple everyday contexts. The file does not provide exclusion conditions or narrower activation guidance to distinguish intentional routing from normal conversation.
Several routes embed English and Chinese trigger patterns directly, but the file does not explain whether multilingual activation is intentional, configurable, or user-selectable. If organizational policy requires user opt-in for language or locale handling, this undocumented hard-coding may violate that expectation.