The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/transcribe.sh suffers from a shell argument injection vulnerability. The command uses unquoted expansions for optional parameters: ${language:+-F "language=${language}"} and ${prompt:+-F "prompt=${prompt}"}. Because these are not quoted, the shell performs word-splitting on the result of the expansion, allowing an attacker to inject additional flags (such as -o to overwrite files or -F to exfiltrate arbitrary files) into the curl execution. Remediation: Use shell arrays to build command arguments or ensure expansions are fully quoted.
[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection as it processes untrusted audio files and user-provided transcription hints. 1. Ingestion points: Raw audio file content and the --prompt CLI argument in scripts/transcribe.sh. 2. Boundary markers: None present to isolate untrusted input from instructions. 3. Capability inventory: Network exfiltration capability via curl and filesystem write capability via shell redirection. 4. Sanitization: No validation or sanitization is performed on the input before it is passed to the shell command or the external API. Remediation: Sanitize inputs and implement explicit delimiters for untrusted data.

openai-whisper-api