The Agent Skills Directory

[COMMAND_EXECUTION]: The subagents for SimpleCov and RubyCritic are instructed to perform shell operations including bundle install and bundle exec rspec. The skill's allowed-tools frontmatter restricts the agent to Read, Grep, and Glob, which contradicts the shell-based execution requirements of the subagents.
[DYNAMIC_EXECUTION]: The skill dynamically modifies executable test helper files (e.g., spec/rails_helper.rb) to inject code for coverage collection. This runtime modification of project source files represents a dynamic execution risk, even with the included cleanup routines.
[INDIRECT_PROMPT_INJECTION]: By auditing untrusted codebases and having the ability to execute commands and write to files, the skill is vulnerable to indirect prompt injection. Malicious instructions embedded in the analyzed code could potentially exploit the agent's capabilities.
[METADATA_POISONING]: The skill provides conflicting author information: 'OpenAI' in the YAML metadata, 'rolemodel' in the developer context, and 'Jose Blanco and thoughtbot, inc.' in the LICENSE file. This discrepancy is a metadata poisoning concern that could mislead users about the skill's source.

rails-audit