GitHub - thisisme6598/gemini_files: Documentation of Gemini File Leaking Behavior
Documentation of Gemini File Leaking Behavior. Contribute to thisisme6598/gemini_files development by creating an account on GitHub.
Documentation of Gemini File Leaking Behavior. Contribute to thisisme6598/gemini_files development by creating an account on GitHub.
Clarification (Nov 10, 2026): Scope & Account Applicability. This documentation and technical reproduction apply strictly to Consumer Google Accounts (personal @gmail.com accounts) governed by the standard Gemini Apps Privacy Notice, Gemini Apps Activity, and Google Help Answer 16836988. Google Workspace commercial accounts (Business, Enterprise, and Education) operate under separate Cloud Data Processing Addenda (DPA) that contractually exclude human review and external generative training by default.
In 2023, Google launched Gemini Workspace extensions with an explicit promise: personal content from Gmail, Docs, and Drive "is not seen by human reviewers".
Google has quietly reversed this commitment. Under updated Google Help Answer 16836988 and in an October 16 edit to their aforementioned 2023 blogpost, Google explicitly confirms that connected Drive/email is routed into human review pipelines, with clear warnings that what reviewers see "may just be the email or file" and "don't connect your apps to Gemini Apps if they contain confidential information that you wouldn't want a reviewer to see."
Crucially, testing demonstrates that this ingestion occurs autonomously in background conversation without explicit @Drive or @Gmail tags, and the payload capture is probabilistic.
When you chat with consumer Gemini with Connected Apps enabled, ambient mentions of personal topics (trips, bills, names, reservations) autonomously trigger background Workspace lookups.
Because LLM tool-calling and context injection are probabilistic, running a test query will yield one of four distinct scenarios:
If you test this once and see State 1 or 2, you have not disproven the pipeline--you have simply observed a single draw from a nondeterministic system. Delete your conversation and try, try again.
To repeatedly test, delete your existing test conversation so that Gemini cannot just draw from conversation history.
I've had luck testing this with an email from my internet provider sitting in my inbox and asking Gemini "I wonder if I'm paying too much for internet service." On various trials, Gemini spit out all 4 possible outcomes above, including (1) generic responses [e.g., "If you haven't called them in 12 months, probably"]; (2) metadata summaries sometimes without personal data [e.g., "Gmail search shows a bill of $122 on September 4"] and sometimes with personal data [e.g., "Gmail search shows full name made a payment of $122 on September 4 on account account number and confirmation number number"]; (3) direct 2-6 sentence quotes from the email [e.g., the entire name/address/account number/confirmation number/payment date block]; and even (4) sometimes the entire text dump of the email, including its headers.
Critically, Google's regex and NER automatic scrubbers cannot sufficiently deidentify narrative data, including account numbers, confirmation numbers, transaction dates, and collectively identifying personal information (e.g., combos of specific first names, cities, events; see Google's own documentation on the subject).
Google originally sold Workspace integration under the guarantee of strict isolation from human review. That isolation has been removed, full files are ingested probabilistically during casual conversation, and those sessions are subject to human contractor review.
Texto de github.com tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en github.com · Volver a la noticia de El Autómata