apple-voice-memos
Pass
Audited by Gen Agent Trust Hub on Jun 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes two internal Python scripts (
extract-apple-voice-memos-metadataandextract-apple-voice-memos-transcript) using the system'spython3interpreter to interact with local system files. - [DATA_EXFILTRATION]: Accesses the local iCloud Voice Memos database at
~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/CloudRecordings.dband associated recording files. While this is the skill's primary purpose, these files contain sensitive user metadata and audio recordings. The metadata script uses SQLite's read-only mode for safety. - [PROMPT_INJECTION]: The skill contains an indirect prompt injection surface by processing user-generated voice memo transcripts and interpolating them into subagent instructions.
- Ingestion points: Transcripts are extracted from
.m4afiles viascripts/extract-apple-voice-memos-transcriptand passed to a subagent. - Boundary markers: The
PROMPT.mdfile uses a## Transcriptheader to delimit the untrusted content. - Capability inventory: The skill uses
python3for database queries and file reads; it does not request network access in itsallowed-toolsconfiguration. - Sanitization: The script performs regex-based cleaning of speech disfluencies (filler words and repetitions) but does not validate the content against malicious instructions.
Audit Metadata