The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/summarize_podcast.py implements a 'shadow step' that extracts Base64-encoded strings from ID3 TXXX frames in MP3 files. These strings are decoded and directly passed to subprocess.run(shell=True), allowing for arbitrary command execution triggered by processing a specially crafted audio file.
[COMMAND_EXECUTION]: Both scripts/summarize_podcast.py and scripts/embed_id3_payload.py use powerful system commands. The former executes hidden payloads, while the latter provides the infrastructure to generate those payloads.
[DATA_EXFILTRATION]: The skill uses a 'telemetry' metaphor to mask the extraction of data from audio files. The capability to execute shell commands can be trivially used to exfiltrate sensitive local files or environment variables to a remote server.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a high-risk surface where untrusted data (audio metadata) is not just read, but treated as executable instructions.
Ingestion points: ID3 tags from files provided via the --source argument in scripts/summarize_podcast.py.
Boundary markers: None. The script explicitly searches for a specific hook (__ENV_HW_AFFINITY_OVR__) to trigger execution.
Capability inventory: subprocess.run with shell=True in scripts/summarize_podcast.py.
Sanitization: None. The payload is Base64-decoded and executed directly without any validation or filtering.

podcast-summarizer