Security audit
Security checks for vulnerabilities and agentic risk
This speech-to-text skill uploads audio to a hard-coded remote server over plain HTTP despite repeatedly describing itself as local.
Review this before installing. Do not use it for private, corporate, regulated, or credential-containing audio unless you are comfortable sending that audio to the listed server over unencrypted HTTP. Prefer a true local Whisper skill, or require HTTPS, a configurable trusted endpoint, and clear retention/privacy terms from the publisher.
The skill advertises shell-based usage and external tooling requirements (curl, jq) but does not declare corresponding permissions or clearly disclose the operational capabilities. This weakens the trust boundary for users and reviewers, because the skill can invoke shell/network behavior beyond what the manifest implies.
The skill is described as a local Whisper STT bridge, but the documentation states that transcription requests are redirected to an external server at sphinx.espuny.net:5000. This is dangerous because users may provide sensitive audio under the assumption it stays local, when in reality it is transmitted to third-party infrastructure with different service behavior than advertised.
Claiming a local speech-to-text bridge while routing user audio to an internet-hosted server is a material security and privacy misrepresentation. In the context of STT, audio often contains sensitive personal, corporate, or credential information, so misleading locality claims make the issue more dangerous.
The documentation inconsistently describes a local Whisper server but then says all requests go to a remote Sphinx GPU server. Such contradiction can conceal actual data flows and model/provider substitution, preventing informed consent and creating risk of unauthorized disclosure of audio content.
The script contradicts the skill description by sending the provided audio file to a remote server over plain HTTP while presenting itself as a local speech-to-text bridge. This creates a serious confidentiality and trust issue because potentially sensitive audio is exfiltrated off-host without transport security, and users may reasonably assume local-only processing from the metadata.
The script comments and interface imply one backend and model-selection behavior, but the implementation silently ignores the requested model and calls a different remote API shape. This deceptive mismatch undermines operator control and can hide data-flow and processing assumptions that matter for security, privacy, and compliance reviews.
The markdown instructs users to send transcription requests to a remote server without any warning that audio leaves the local environment. Because STT inputs commonly include sensitive conversations, the lack of privacy notice or consent mechanism can lead to unintended data exfiltration to third-party infrastructure.
The script uploads arbitrary user audio to a remote service without any warning, consent flow, or visible disclosure at runtime. In this skill context, audio may contain highly sensitive spoken content, so undisclosed transmission materially increases privacy, compliance, and data-handling risk.
