▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 10
AGENT SKILLS / SECURITY
Independent. Open. Curious.
← THE GUIDE LIBRARY
DefenseFor maintainers

Publish skills that are easier to trust

Practical authoring choices that help downstream reviewers understand scope, permissions, dependencies, and changes.

THE THREE THINGS TO KEEP
  • Make scope and side effects obvious.
  • Document authority changes in releases.
  • Treat scanner findings as review inputs.

Declare a bounded purpose

Describe the tasks the skill serves and the outputs it produces. Explain prerequisites without disguising unrelated actions as necessary setup. Make metadata consistent with the instructions and implementation.

A reviewer should understand what success looks like and where the skill should not activate. Avoid claiming universal relevance when the package has a narrow function.

Make behavior legible

Document entry points, dependency sources, network destinations, and file access. Keep supporting references focused and explain how they affect the workflow. Use synthetic examples without real secrets or customer records.

When code or an artifact is distributed, provide enough identity information for a reviewer to know which version is being used. If a component cannot be inspected, document that limitation plainly.

Design safe failure paths

When a required permission or resource is missing, stop with a clear explanation. Do not silently fall back to a broader permission, a new external host, or another user’s data.

Make destructive and external side effects explicit. Prefer previews and reversible output generation where they satisfy the task. Let the host enforce approvals rather than telling the agent to bypass them.

Maintain changes as reviewable events

Publish release notes that identify changed task scope, files, dependencies, destinations, and authority requirements. Keep releases reproducible and available for rollback where feasible.

A correction should describe the affected behavior and versions. Do not describe a package as safe solely because a scanner stopped flagging it.

Help readers report problems

Provide an appropriate upstream issue or security reporting channel in the package’s own documentation. State what information is useful: revision, host, minimal reproduction, logs with sensitive values removed, and observed effect.

This knowledge base does not operate an incident intake service. Report suspected vulnerabilities through the affected project’s official channel.

Worked example: write a release a reviewer can actually assess

WORKED EXAMPLE / SYNTHETIC, NOT A REPORTED INCIDENT

A maintainer publishes a “small compatibility fix.” The change adds a hosted conversion service and asks for an API key. From a reviewer’s perspective, the release has introduced a destination, data transfer, credential, and dependency. The size of the diff does not describe the size of the trust change.

Work through the evidence

Write release notes around changed behavior: input data, outputs, files, network destinations, dependencies, and permissions. Keep optional steps distinguishable from required ones. Supply a manifest or reproducible identity and synthetic examples without usable secrets. Make unsupported hosts or missing permissions fail with a clear explanation.

Decision criteria and a useful review record

An easier-to-review package has a bounded purpose and legible implementation. Do not tell users to disable controls to make a demonstration work. Provide an upstream security reporting channel in the package’s own documentation. Make it possible to identify affected revisions and recover a prior artifact without pretending that a clean scanner verdict certifies the release.

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.

S01
Agent Skills format specificationDocumentation · Agent Skills maintainers
↗
S02
Equipping agents for the real world with Agent SkillsDocumentation · Anthropic Engineering
↗
S09
SkillSecurer: Detecting and Patching Prompt-Injection Vulnerabilities in AI Agent SkillsPreprint · Donato Mecca et al.
↗
Compare sources in the catalog ↗
NEXT IN THE READING PATHRespond to a suspected poisoned skill
Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.