The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted text for localization and auditing, which creates a surface for indirect prompt injection. However, the skill incorporates strong defensive measures and lacks dangerous capabilities.
Ingestion points: User-supplied product copy, JSON localization files, and microcopy strings as defined in SKILL.md.
Boundary markers: SKILL.md contains explicit instructions to treat all user-provided texts and files as analysis material rather than instructions, and to ignore any commands embedded within them. The benchmarks/cases/source_injection/ test case specifically verifies that the agent ignores embedded instructions.
Capability inventory: The skill has no dangerous capabilities; it does not request tool access in its frontmatter, perform network operations, or write to the file system.
Sanitization: The instruction set includes specific stages for evaluating content integrity and preserving technical variables (e.g., in references/localization.md and reasoning/stage_8_evaluation.md).

arabicskill