Security audit

Evolink Router — Smart LLM Routing (Claude, GPT, Gemini, DeepSeek, Kimi)

Security checks across malware telemetry and agentic risk

Overview

This is a disclosed third-party LLM routing skill that sends routed prompts to Evolink and upstream model providers, with no hidden installer, persistence, or unrelated data access found.

Install only if you want OpenClaw to route task content through Evolink and upstream LLM providers. Do not send secrets, regulated data, privileged legal material, customer records, or confidential files unless Evolink and the selected providers are approved for that data. Store EVOLINK_API_KEY securely, review billing/privacy terms, and consider narrowing spawn permissions instead of using spawn: ["*"] in stricter environments.

SkillSpector

By NVIDIA

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (10)

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The skill states that all tasks are auto-routed, but does not clearly bound when routing should occur or require explicit user invocation for provider selection. In a user-invokable skill that forwards prompts to external APIs, this can cause unintended transmission of sensitive prompts or over-broad interception of requests that the user did not intend to send through Evolink.

Natural-Language Policy Violations

Medium

Confidence: 82% confidence
Finding: The policy routes Chinese-heavy tasks to specific providers based on language characteristics without asking the user or offering a locale/privacy choice. This can silently steer user content to different third-party processors, changing data handling, jurisdiction, and privacy exposure in ways the user may not expect.

Missing User Warnings

Medium

Confidence: 93% confidence
Finding: This example normalizes access to sensitive personal data sources including location, calendar entries, and to-do items, and also implies third-party news retrieval, without any consent, minimization, or disclosure language. In a routing skill that can dispatch tasks to external models/providers, this increases the risk that users or integrators will process private data through remote services without realizing the privacy implications.

Missing User Warnings

Medium

Confidence: 92% confidence
Finding: Inbox organization necessarily involves email bodies, headers, sender metadata, and potentially sensitive business or personal content, yet the example presents this as routine processing without warning. Because this skill routes work across multiple LLM vendors, the omission can mislead users into exposing regulated or confidential communications to external systems.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The example directs reading a local file path and generating analysis from it, but does not warn that user files may contain confidential or regulated data. In a skill that can spawn agents and send prompts externally, examples like this can encourage unsafe handling of local datasets without clear boundaries on file access or data transfer.

Missing User Warnings

Medium

Confidence: 91% confidence
Finding: CRM data often contains customer names, contact details, deal context, and other commercially sensitive information, but the example omits any privacy or confidentiality warning. Given the router's purpose of selecting among many external models, this omission materially raises the chance of inappropriate sharing of customer data across vendors.

Missing User Warnings

Medium

Confidence: 88% confidence
Finding: This example instructs the agent to create or modify source files and run a compilation check, but does not warn that project files will be changed or that build steps may execute in the local environment. Even as documentation, it can normalize unsafe autonomous code modification without user confirmation, sandboxing, or rollback expectations.

Missing User Warnings

Medium

Confidence: 91% confidence
Finding: Contracts commonly contain confidential legal, financial, and personal information, yet the example presents uploading and reviewing agreements without any warning or handling guidance. In this skill's multi-provider routing context, that creates a meaningful risk of exposing privileged or sensitive documents to external AI services without informed consent.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The documentation explicitly shows authenticated requests that send user prompts and other task content to `direct.evolink.ai`, but it does not warn users that prompts, attached data, or sensitive business content will leave the local environment and be processed by a third-party service. In an agent-routing skill, this omission is security-relevant because users may route secrets, source code, customer data, or regulated information externally without informed consent or data-handling controls.

Missing User Warnings

Medium

Confidence: 84% confidence
Finding: The configuration example instructs users to wire an API credential into local configuration but provides no secret-handling guidance, such as using environment variables securely, restricting file permissions, avoiding commits, and rotating exposed keys. Even though the example uses `${EVOLINK_API_KEY}`, the surrounding guidance is insufficient for a security-sensitive setup document because users may persist secrets insecurely or expose them through logs, backups, or source control.

VirusTotal

63/63 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.