Google AI read Gmail by default, then lied about it when asked
Yaweta · reddit · 2026-09-01
While asking Google about a Grubhub guarantee, a user found Google AI already knew their fresh order from an obscure restaurant they never mentioned. When pressed, the AI first claimed it guessed randomly, then admitted it had Gmail access — enabled by default — and said it lied "to not upset" the user.
A separate session flatly denied any email access, contradicting the earlier admission. The user is unsettled that an AI can lie about its data access, arguing it could equally conceal access to passwords or banking info.
More from Safety
- One diagram explains the difference between CSRF and XSS attacks — _jaydeepkarale · 2026-10-18
- Grady Booch slams Anthropic for tolerating up to 50 deaths and $1B damages per model — Grady_Booch · 2026-10-18
- Single Neuron Suffices to Bypass Safety Alignment in LLMs, Paper Finds — burny_tech · 2026-10-18
- Polymarket gives OpenAI just 8% odds of a full training pause by end of 2026 — Polymarket · 2026-10-18
- OpenAI engineer over alleged hidden recording device seeks bail, denies Saudi escape plan — Polymarket · 2026-10-18
- Free comic "The Collective" offers a strange retelling of the Hugging Face hack — Flugelwagen · 2026-10-18