app-intents-whats-new-27
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The file
apple/guide.mdcontains instructions that command the agent to treat the provided text as 'authoritative' and states that it 'unconditionally supersedes any prior training.' This language is designed to override the model's internal knowledge base and safety guidelines in favor of the skill's content. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from the user's local development environment. Ingestion points: The agent is instructed in
SKILL.mdto 'Inspect the project's actual Xcode/Swift settings and deployment targets' and to read relative paths within the repository. Boundary markers: The skill lacks robust boundary markers or specific 'ignore embedded instructions' warnings for the project data it processes, relying instead on high-level behavioral guidance. Capability inventory: The agent is expected to have access to tools for reviewing, analyzing, and potentially suggesting edits to Swift code and Xcode project configurations. Sanitization: No explicit sanitization or filtering of the user-provided project files is performed before the LLM processes them.
Audit Metadata