Repository navigation
Add a security rule for creating agent skills #110
Description
Summary
Add a model-agnostic CodeGuard rule for securely creating and modifying agent skills.
Skills can control tools, files, networks, credentials, and agent behavior, so they should be treated as security-sensitive artifacts.
Related discussion: #43
Proposed guidance
The rule should cover:
- Using only necessary tools and permissions.
- Making filesystem, network, shell, and credential access explicit.
- Avoiding hidden instructions and unsafe commands.
- Documenting the skill's purpose, inputs, outputs, and dependencies.
- Reviewing skills before publishing or installing them.
Acceptance criteria
- A new source rule is added.
- Relevant examples are included.
- Validation and existing tests pass.
- Generated formats and documentation are updated as needed.
Activity
AdvaitVarhade commentedon Aug 4, 2026
I would like to work on this. I can implement the new security rule in Python to analyze skill definitions for tool permissions, unsafe shell commands, and hidden instructions, along with complete test cases and documentation updates.
thomas-bartlett commentedon Aug 31, 2026
Thank you @AdvaitVarhade for the work on PR #112, and apologies for the delayed follow-up.
Before moving forward, we would like to compare the proposal with existing CodeGuard guidance and current approaches to securing agent skills so we can identify the concrete gap and the right artifact. No action is needed while we complete that review.
We appreciate the contribution and your patience.