The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates high security awareness by explicitly instructing the agent to avoid direct interpolation of untrusted event data into AI prompts. It mandates using reference properties and escaping data to prevent forged tags or instruction hijacking from Slack messages or GitHub issues.
[PRIVILEGE_ESCALATION]: The skill implements a principle of least privilege by defaulting to 'read_only' scopes for MCP connections and requiring explicit user authorization for 'full' access. It also enforces 'actor_access' filters on GitHub triggers to ensure only authorized users can initiate tasks.
[COMMAND_EXECUTION]: The workflow creation process uses well-defined internal templates and tools (e.g., template-posthog-create-task). It requires verifying the reachability of repositories and channels using official integration tools rather than relying on unverified user input.
[REMOTE_CODE_EXECUTION]: While the skill builds workflows that execute AI tasks, it relies on a managed environment with predefined connectors and skills, reducing the risk of arbitrary remote code execution.

building-loops