The Agent Skills Directory

[METADATA_POISONING]: The skill describes fictitious software versions (iOS 26+) and APIs (SpeechAnalyzer, SpeechTranscriber, AssetInventory) that do not exist in the current Apple ecosystem. It directs users to a third-party domain (sosumi.ai) for documentation rather than official Apple developer resources.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio data from the microphone or files into text, creating a vulnerability surface where transcribed instructions could influence downstream agent behavior. * Ingestion points: Microphone audio (via AVAudioEngine) and audio files (via AVAudioFile) as described in SKILL.md. * Boundary markers: Absent; no instructions or code mechanisms are provided to distinguish between transcribed data and control instructions. * Capability inventory: Accesses microphone audio and file system to perform speech-to-text transcription using the Speech framework. * Sanitization: No validation, filtering, or sanitization logic is applied to the resulting transcription strings before they are used or displayed.

speech-recognition