SkillClone: Multi-Modal Clone Detection and Clone Propagation Analysis in the Agent Skill Ecosystem
This paper introduces SkillClone, the first multi-modal clone detection system for agent skills that fuses TF-IDF and per-channel analysis to identify widespread duplication and propagation risks in the ecosystem, revealing that the current landscape is significantly inflated with redundant and often superseded variants.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a massive, bustling library where millions of people have written "instruction manuals" for robots. These manuals, called Agent Skills, tell the robots how to do specific tasks like writing code, checking for security holes, or organizing files.
Recently, this library exploded in size. There are now nearly 200,000 of these manuals. But there's a huge problem: Nobody knows who copied whom.
The Problem: The "Silent Copycat" Epidemic
In the old days of software, if you copied a piece of code, you usually had to declare it or use a specific package manager that tracked dependencies. But in this new world of Agent Skills, developers just copy-paste a manual, tweak it a little, and upload it as their own.
This creates a systemic risk.
- The Analogy: Imagine a chef writes a recipe for a soup that accidentally contains a deadly poison.
- The Reality: If 1,000 other chefs copy that recipe, change the font, rename the dish, and swap out the vegetables but keep the poison, the poison spreads silently.
- The Danger: If a security researcher finds the poison in the original recipe, they can't warn the other 1,000 chefs because the library has no way to link the copies back to the source. The poison keeps spreading, invisible to everyone.
The Solution: SkillClone (The "DNA Scanner")
The authors of this paper built a tool called SkillClone. Think of it as a super-powered DNA scanner for these instruction manuals.
Most existing tools are like single-lens cameras. They only look at one thing:
- Code-only scanners: They only read the computer code. If the code is rewritten, they think it's a new, unique manual.
- Text-only scanners: They only read the English instructions. If the words are changed, they miss the connection.
SkillClone is a 3D holographic scanner. It looks at three different "channels" of the manual simultaneously:
- The YAML Header: The metadata (title, tags, settings).
- The Natural Language: The English instructions explaining how to use the skill.
- The Code Blocks: The actual computer code that does the work.
Even if a copycat changes the code completely but keeps the title and the instructions 90% the same, SkillClone spots the connection. It fuses these three clues together to say, "Hey, this isn't a new idea; this is a clone of that other idea."
How It Works (The "Detective" Logic)
The tool uses a clever trick. It doesn't just add up the similarities; it looks for patterns.
- Scenario A: Two manuals have identical titles, identical instructions, and identical code. -> Type 1 Clone (Exact Copy).
- Scenario B: Two manuals have identical titles, but the instructions are rewritten, and the code is totally different. -> Type 4 Clone (Same idea, totally different execution).
SkillClone is the first tool that can catch Scenario B. It realizes that even though the code is gone, the "soul" of the manual (the title and the instructions) is still a perfect match.
What They Discovered (The Shocking Stats)
When the researchers ran SkillClone on 20,000 skills, they found some eye-opening truths:
The Library is 3.5x Bigger Than It Looks:
If you remove all the copies, the 20,000 skills shrink down to just 5,642 unique ideas. The rest are just re-skinned versions of the same thing. The ecosystem is "inflated" with duplicates.Most "New" Skills are Just "Remixes":
75% of all skills are part of a "clone family." Most people aren't writing from scratch; they are taking an existing skill, changing a few words, and calling it their own.The "Super-Spreaders":
A tiny group of authors (less than 1%) created over half the skills. Worse, some "aggregator" bots are just copying thousands of skills from one author and re-uploading them as their own, creating a fake illusion of diversity.The Security Nightmare:
They found 141 skills that contained real, dangerous code (like SQL injection or reverse shells).- Because of the cloning, these dangerous patterns didn't stay in one place.
- They spread to 1,100 other skills.
- Crucially, 71% of these dangerous clones were made by different authors who had no idea they were copying a virus. They thought they were just improving a tool.
Why This Matters
Before this tool, security teams were like firefighters trying to put out a fire by only looking at the house they were standing in. They didn't know the fire had already jumped to the houses next door, and the houses across the street.
SkillClone gives them a map of the entire neighborhood. It allows them to:
- Trace the Blast Radius: If one skill is dangerous, they can instantly find all 1,000 clones that need to be fixed.
- Clean Up the Library: They can identify the "best" version of a skill and hide the 50 inferior copies, making the library smaller and safer.
- Stop the Silent Spread: They can alert maintainers that their "original" work was actually copied from a vulnerable source.
In short, SkillClone is the first tool to bring order to the chaos of the Agent Skill ecosystem, turning a wild, untracked jungle of code into a manageable, secure library.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.