The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to fetch component source code from an external, non-whitelisted domain. In signature-effects/SKILL.md, instructions specify to 'Fetch the source from the registry (https://ui.aceternity.com/registry/.json)' and integrate the retrieved code into the local project. This introduces a risk of incorporating unverified or potentially malicious remote code into the user's application.
[COMMAND_EXECUTION]: The skill includes and instructs the agent to execute local diagnostic scripts. interface-guidelines/SKILL.md directs the agent to run a shell script using bash <this-skill-dir>/scripts/scan.sh, and devtool-visual-system/SKILL.md directs the use of a Python script via python3 <this-skill-dir>/scripts/theme.py. While these specific scripts appear to be diagnostic, they establish a pattern of executing local files distributed with the skill.
[PROMPT_INJECTION]: The main router skill (SKILL.md) contains aggressive directives aimed at overriding standard agent behavior and restraint. The agent is instructed to 'show no restraint,' stated to have 'full authority to throw the current design out,' and told 'don't ask permission to discard what isn't working.' These instructions are designed to bypass the agent's typical safety protocols regarding caution and user confirmation for significant destructive actions.
[INDIRECT_PROMPT_INJECTION]: The skill's auditing functionality creates an attack surface for indirect injection. Ingestion points: The agent reads arbitrary project source files during the UI audit process. Boundary markers: There are no explicit markers or instructions for the agent to ignore directives embedded within the ingested files. Capability inventory: The agent has the capability to execute shell/Python scripts and write new code to the filesystem. Sanitization: No sanitization or validation is performed on the ingested code before it is processed by the LLM, potentially allowing malicious project files to influence agent behavior.

bespoke