What an AI Agent Pulls In With a Single Skill
mlj.solutions
mlj.solutions Studie ukazuje, že u AI skills zůstávají verze, původ a závislosti většinou implicitní, což už dnes vede k duplicitám a nekonzistentním instalacím napříč komunitou. Autoři proto navrhli SkillDepAnalyzer – nástroj inspirovaný SBOM (Software Bill of Materials), který navíc zachycuje i závislosti popsané přirozeným jazykem, ne jen strojové manifesty. Na benchmarku SKILL-DEP je přesnější než LLM baseline i klasické SBOM nástroje. 🔗 Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains ❓Jak řešíte evidenci dependency grafu u vašich AI skills?
View on Instagrammlj.solutions
mlj.solutions Nainstaluješ bezpečně vypadající AI skill. Riziko se skrývá o tři závislosti dál. Přesně proto je čím dál důležitější hlídat si, na čem tvůj skill (nebo agent) vlastně závisí a ne jen to, jak vypadá na první pohled.
View on InstagramWhat the study is about
A skill is a package of knowledge and procedures that an AI agent uses while working. But skills rely on other skills, packages and services, and nobody tracks their versions or origin. The study mapped this across 1.43 million skills and shows that checking just one skill isn't enough.
Statistics
1.43M
skills analyzed
This is a large-scale analysis from which the authors built a map of dependencies between skills, packages and services.
4
structural patterns
The authors found four recurring patterns in the data, such as concentrated reuse or recursive growth of dependencies.
3
types of dependencies
Dependency graphs link skills, software packages and external services, often invisibly.
What this means
- Dependencies are invisible, the identity, version and origin of skills remain unclear, which leads to duplicates and inconsistent installations.
- Checking the skill itself isn't enough, security signals are often hidden only in its dependencies.
- Reuse quietly inflates the inventory, skills that call other skills bring packages into the system that nobody knows about.
- Malicious skills persist, the authors found known malicious skills in the ecosystem and reported them to their developers.
- A solution exists, the authors recommend typed dependency manifests and exact-version records for skill developers.
Our take
An AI agent is only as secure as the weakest thing it pulls in. That's why, with corporate agents, we check not just the skills themselves but everything they are built on.
Zdroj
Jia, C., Zhao, T., He, R., & Zhou, M. (2026). Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains. arXiv:2607.01136. https://arxiv.org/abs/2607.01136
Don't miss these posts
When More Agents Pay Off and When They Don't
More AI agents don't necessarily mean a better result. A study shows that what matters is how much information is lost when it is passed between agents.
AI Remembers a Claim Even When You Tell It It's False
When a model is fine-tuned on documents that deny a claim, it often learns the claim as true anyway. The authors named this effect Negation Neglect.
The Skill That Explains Anything to Anyone
ELI5 is an open-source skill for Claude Code that adapts explanations to the audience. With the skill it passed 10 of 12 criteria, without it only 5.
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
