The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill mandates a mandatory sandbox bypass, requiring all invocations of the Cursor CLI to request escalated/host permissions on the first attempt. It explicitly forbids the use of the default restricted sandbox and instructs the agent to report the task as blocked if escalation is denied.
[COMMAND_EXECUTION]: The skill provides instructions for executing external binaries (agent, cursor-agent, cursor) via the host shell with parameters like -p and --model.
[INDIRECT_PROMPT_INJECTION]: The skill implements a delegation pattern that is vulnerable to indirect prompt injection. 1. Ingestion points: Untrusted data enters context when the Calling Agent builds a prompt for the Cursor CLI using user-provided objectives and repository content. 2. Boundary markers: The skill suggests a structured contract but lacks technical delimiters to prevent the secondary agent from obeying instructions embedded in code. 3. Capability inventory: The delegated Cursor tool has broad capabilities including file search and multi-file refactoring. 4. Sanitization: No specified sanitization for content interpolated into the Cursor prompt.
[METADATA_POISONING]: The skill's YAML frontmatter contains a future date ('2026-09-21') in the 'updated' field, which is misleading regarding its provenance.

cursor-delegate