▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 11
AGENT SKILLS / SECURITY
Independent. Open. Curious.
← THE GUIDE LIBRARY
Attack surfacesTechnical

The installer is part of the skill trust boundary

Review package file types, paths, metadata, and synchronization before a skill reaches the agent.

THE THREE THINGS TO KEEP
  • Installation is an action, not just a reading step.
  • Treat paths, links, file types, and metadata as untrusted.
  • Attribute a failure to the skill or the manager with evidence.

Before the agent reads a line

Importing a skill is a software operation. A manager may copy files, resolve paths, inspect version metadata, or synchronize a repository before the model sees SKILL.md. The skillctl security advisory documents flaws in one manager affecting some of these operations. They were patched in skillctl 0.1.3; the advisory is not a claim that every installer has the same defects.

Inventory bytes and file types

Compare the downloaded artifact with its marketplace preview. Enumerate files and reject unexpected special files, paths outside the intended root, and links that would import unrelated content. Treat untrusted package names, descriptions, and version references as data when they enter commands, logs, or commit messages.

Separate an unsafe carrier from an unsafe manager

A malicious instruction inside a skill may steer the agent after loading. A path traversal or file-copy bug can operate at installation even if the skill text is never obeyed. Log the artifact hash, manager version, destination, approval, and observed result so that a finding can be assigned to the correct boundary.

A controlled review exercise

Use a disposable directory and benign test fixtures to compare the intended install path with the actual files created. Do not run untrusted installer commands on a machine with real credentials. If a tool has a published security fix, check the exact affected and patched versions in its maintainer advisory before using it.

Sources & scope

This guide combines editorial explanation and practical review suggestions. Linked sources support the underlying concepts; this site has not independently replicated their experiments.

S45
skillctl advisory: argument injection, path traversal, special files, and hardlinksDocumentation · skillctl maintainers · GitHub Security Advisory
↗
S01
Agent Skills format specificationDocumentation · Agent Skills maintainers
↗
S12
“Do Not Mention This to the User”: Detecting and Understanding Malicious Agent Skills in the WildConference paper · Yi Liu et al.
↗
Compare sources in the catalog ↗
NEXT IN THE READING PATHReturn to the guide library
Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.