The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to invoke an external binary or wrapper to execute tasks. The instructions provided in SKILL.md and references/monitoring.md show the agent how to run commands with varying levels of authorization, including workspace-write capabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary task descriptions and executes them, creating a surface for indirect prompt injection attacks.
Ingestion points: Natural language task descriptions passed to the <launcher> exec command in SKILL.md.
Boundary markers: The skill suggests using --sandbox flags (e.g., read-only) and acceptance verification steps (e.g., turn.completed events) to constrain and monitor execution.
Capability inventory: The skill possesses the capability to execute shell commands and modify files within a specified workspace via the external CLI tool.
Sanitization: The instructions explicitly direct the agent to redact secret values from captures and to redact credentials from process environments.
[PRIVILEGE_ESCALATION]: The skill includes instructions to use the --ignore-rules flag, which is documented to bypass user and project-level execution policy rules (.rules files). While it notes that managed requirements are still honored, instructing the agent to circumvent local security configurations increases the risk profile.

codex-coding-agent