Persona Researcher Skill Security Audit

These are live instructions, not examples or warnings: the skill directs the agent to announce files to collaborators and request reviews through Gmail. The visible instructions do not require prior confirmation of recipients, message content, link permissions, or whether the research may be disclosed. If executed with an authorized account, material could therefore reach unintended recipients or be shared too broadly. Users can require send-time previews, explicit recipient approval, and least-privilege link settings.

These are active cloud-modification instructions: organize Drive content, write Docs, and append to Sheets. The visible text does not identify target files or folders, define editing boundaries, or require previews or confirmation; if the agent selects targets itself, it could alter the wrong material or contaminate shared records. The tip to keep a running shared-sheet log also supports repeated writes. Users can restrict operations to explicit file IDs and require a preview or copy before modification.

The manifest and prerequisite do require Drive, Docs, Sheets, and Gmail skills, and the later instructions span discovery, modification, and outbound communication. However, the supplied source does not show their permission scopes, authorization flow, or implementations, and loading a skill does not necessarily grant every Google account permission at once. The capability bundle is broad, but actual over-privileging cannot be established from these lines alone. Users can ask for the OAuth scopes and whether Gmail and write access can be authorized separately per task.