smart-glasses-live-captions
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a method for generating 'short hints' from live speech using an LLM. As documented in
references/short-hints.md, the prompt template interpolates untrusted speech-to-text data directly into the model's context. \n - Ingestion points:
references/short-hints.md(via theInput:field in the prompt template).\n - Boundary markers: Absent; the template uses simple text headers without robust delimiters to isolate untrusted speech content.\n
- Capability inventory: The resulting text is displayed on smart glasses and can be spoken via text-to-speech engines.\n
- Sanitization: Absent; there is no documented logic for filtering or escaping malicious instructions embedded in the captured speech.\n- [EXTERNAL_DOWNLOADS]: The installation instructions for the example app in
SKILL.mdincludenpm installcommands, which download third-party dependencies from the NPM registry.\n- [COMMAND_EXECUTION]: The skill documentation provides several shell commands for setting up a Node.js server and running hardware simulators, which require the user to execute commands on their local machine.
Audit Metadata