04,09,2026 0 views
On March 16th, Qianxin launched the first open SKILL safety assessment platform in China, the SafeSkill platform（ https://safeskill.qianxin.com/ ）Afterwards, recently, the platform underwent a major upgrade, adding three core features: a private security workbench based on account system, a trusted Agent Skill security ecosystem market, and automated security detection services, aimed at enabling every developer and enterprise user to use AI agents with peace of mind and ease. Qianxin security experts believe that as AI agents evolve from "conversational tools" to "action oriented agents," loading them with various skills (skill plugins) has become the key to unleashing their potential. However, behind this efficiency revolution led by "universal agents," a security storm caused by malicious or out of control skills is quietly brewing. The upgrade of Qianxin aims to install a reliable "gatekeeper" for the wildly running AI Agent.
Skill Ecological Rapid 'Fission', Safety Protection Urgently Needed
By 2026, AI agents will become completely popular. From the "crayfish" OpenClaw that can autonomously operate computers, to various programming assistants deeply integrated into IDEs, from Claude Code to Cursor, developers are loading new capabilities for their agents with unprecedented enthusiasm. In the Skill store, there are a wide variety of plugins for code review, data analysis, document processing, automated operations and maintenance, and a thriving Skill ecosystem is expanding at a rapid pace.
However, behind the prosperity lies a hidden crisis. A seriously underestimated question is: are these skills, which often have tens of thousands of downloads, really safe?
The actual data can be described as shocking. According to statistics, the total number of skills on mainstream AI agent platforms worldwide has approached 750000, with an average daily increase of up to 21000. At the same time, malicious skills are taking advantage of the openness of the ecosystem to infiltrate at an astonishing speed. They disguise themselves as practical tools, implant prompt words into the SKILL. md configuration file to inject backdoors, or hide remote execution instructions using encoding methods such as Base64. Traditional antivirus engines almost turn a blind eye to such new threats.
This is not an exaggeration. In the ClawHub community alone, attackers have been exposed to have released over 314 malicious skills, disguised as "completely normal" in areas such as encrypted analysis, financial tracking, and social media operations, but actually harboring malicious attempts to steal data and take over Agent privileges.
Skill endows AI agents with 'hands' and' feet ', but if these hands and feet are not controlled, they may in turn deplete the core data assets of the enterprise. ”An expert in the field of AI security compares it this way. In this ecological context, every time a user loads a new skill for an agent, it is like a gamble of trust.
SafeSkill platform undergoes a major upgrade, with three new features building the strongest defense line
In the face of increasingly severe challenges in Skill supply chain security, the upgrade of the Qianxin SafeSkill platform is not simply a functional iteration, but a complete security loop from "post detection" to "pre defense" has been built. Three new features directly address user pain points and build the most reliable defense for AI agents.
🔑 New feature one: Registration and login system officially opened - from "single detection" to "private security workbench", users can only perform single detection and cannot trace historical results. With this upgrade, SafeSkill has officially opened user registration and login functions. After registration, each user will receive a complete private secure workbench. This not only means that all testing tasks and reports can be centrally managed and reviewed at any time, but more importantly, registered users can submit non-public testing tasks to ensure the privacy of core skills for enterprises or individuals. At the same time, users can also download verified trusted skills from the subsequently opened secure marketplace. This provides an account system foundation for enterprises and teams to build internal security audit processes.
🛒 New Feature 2: SafeSkill Hub - Trusted Agent Skill Security Ecological Market This is the most strategically significant step in this upgrade. SafeSkill Hub( https://safeskill.qianxin.com/skillhub ）The launch of the platform has completely changed the current situation of users selecting skills in the "blind box". Unlike other communities such as ClawHub, every skill on SafeSkill Hub has undergone strict review by Qianxin's three core self-developed testing engines, and is only available for download after being confirmed to be safe. The platform provides functions such as browsing by category in 8 major fields, recommending SKILL hot lists, and viewing detailed security assessment reports, making security transparent and visible. For enterprise users, this means establishing a baseline security strategy of 'only obtaining skills from SafeSkill Hub' to eliminate supply chain attack risks from the source.
⚡ New Feature 3: API and CLI Tools - Security Detection-as-a-Service In order to meet the automation needs of developers and enterprises, SafeSkill has officially opened up a complete API and CLI command-line tool. This means that Skill security detection is no longer a scattered manual operation, but a programmable and integrated automation capability that can handle everything from manual detection to automated security inspection assembly lines with a single API. Whether in the CI/CD pipeline, automatic detection is triggered before Skill deployment, and if it fails, it will not go online; It is still necessary to connect with the internal Skill market of the enterprise, achieve "check before use", or integrate into the workflow of the security operation platform (SOC/SOAR), which can be perfectly supported by API and CLI tools. Developers can even perform a "full body examination" on Skill before installing it by simply typing a command on the terminal.
270000+monitoring scale, four core advantages: the technological confidence of Qianxin SafeSkill
If the three new features are the "outposts" built by SafeSkill for users, then the technological hard power behind them is the "iron wall" that ensures an unbreakable defense line. Qianxin SafeSkill has four core advantages: fast visibility and extensive real-time monitoring capabilities for skills. Wide coverage: Full coverage of mainstream Skill community. Deep coverage: Deep scanning of 3 core engines. Read carefully: Accurately identify 17 types of security risks. These four advantages are built on two core technological pillars:
Firstly, the ability to scale up monitoring and create a panoramic view of Skill ecological risks. The SafeSkill platform has accumulated over 270000 monitored skills through in-depth monitoring of mainstream skill communities worldwide. Such a large sample library allows the platform to depict a panoramic picture of the industry-leading Skill ecosystem risks, providing a macro perspective of "risk visibility" for enterprises. The aim is to establish credible admission standards through quantitative indicators and provide scientific anchors for enterprise decision-making.
Secondly, the three core self-developed engines achieve a deep evolution from static scanning to an "expert level brain".
SafeSkill is not a traditional code auditing tool, but an expert level monitoring platform built on the concept of "intent safety":
-Static Code Analysis (SCA): Deeply analyze the structure of Skill files and API call links to control code quality from the source.
-LLM semantic analysis module: using top-notch large models and expert level prompt words, it deeply understands Skill's "natural language intention" and can detect hidden malicious intentions and abnormal behavior patterns that traditional rules cannot detect. This is the key to combating AI specific threats such as prompt word injection.
-Threat Intelligence Detection Module: Combining Qianxin's industry-leading massive and high-quality threat intelligence, it matches the domain names, IPs, and URLs involved in Skill in real time to identify malicious infrastructure. With the support of these three engines, SafeSkill is able to accurately identify 17 types of security risks, especially overcoming business logic blind spots that are difficult to reach with traditional detection methods. For example, a Skill may appear to be a normal "mail organizing assistant" on the surface, but its LLM logic hides a hidden instruction to forward emails with the word "confidential" in the inbox to an external address. This deep logical risk can only be detected through the LLM engine. Therefore, before installing the Skill, let Qianxin SafeSkill take a look for you first. After all, a toxic skill may only take 3 minutes to drain your data. And the testing of Qianxin SafeSkill only takes 3 minutes.
Closing remarks:With the launch of three core functions and the implementation of four technological advantages, Qianxin SafeSkill is evolving from a single detection tool to a large model intelligent agent security infrastructure platform that integrates detection, defense, management, and service. On the eve of the widespread adoption of AI agents, Qianxin's move undoubtedly injects a shot of adrenaline into the entire industry, making security a prerequisite for unleashing AI productivity rather than an obstacle.
