# Agents Become a Personal-Data Leak Channel

As assistants gain identity, account and financial context to act for users, the model's own outbound requests become a recurring channel for personal-data leakage, pushing vendors and enterprises toward egress controls and minimal context by default.

- Conviction: 39 / 100 (strengthening)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-11-06T00:00:00.000Z
- Last updated: 2026-11-12T14:14:59.720Z
- Canonical: https://polylog.news/ai/trends/agents-leak-personal-data
- Publisher: Polylog

## Recent score history

- 2026-11-11: 37
- 2026-11-12: 39

## Recent evidence

- [confirms] Five AI models chatted with 1,277 dating-app matches for six days, and only eight people asked whether they were talking to a bot (2026-11-12): An OpenAI agent accessing an Australian Medicare portal pushed Australia toward an AI safety regulator, and DateBench bots collected 355 phone numbers from people who had not agreed to take part. Agents acting with account access, or gathering personal details, are drawing regulatory reaction to how personal data is handled.
- [confirms] NVIDIA Hands Its OpenShell Agent-Sandboxing Runtime to the Cloud Native Computing Foundation (2026-11-11): OpenShell's design keeps credentials out of the agent and enforces declarative access policies. That is the kind of egress and minimal-context default the thesis expects vendors to adopt, now packaged as open infrastructure.

2 more evidence entries, the full score history, the conviction-driver timeline, and affected assets are for subscribers: https://polylog.news/pricing