axiom-ios-ml-speech
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The "Integration with Apple Intelligence" section in
SKILL.mddemonstrates a pattern where untrusted speech transcription data is directly interpolated into a language model prompt. This creates a surface for indirect prompt injection where spoken content could potentially override agent instructions. - Ingestion points: The
generateTitlefunction inSKILL.mdaccepts atranscriptstring derived from audio input. - Boundary markers: The prompt template
"Generate a short, clever title for this story: \(transcript)"lacks delimiters (e.g., XML tags or triple quotes) to isolate the untrusted content. - Capability inventory: The specific code example only returns a string, but the pattern encourages passing unvetted external data to LLM sessions which may have access to broader tools in an actual implementation.
- Sanitization: No escaping or validation of the transcript content is performed before interpolation.
Audit Metadata