Meta
Disclosed Jun 5, 20263 months agoConfirmed
Employee keystroke and screen data collected to train AI was widely exposed inside Meta
Meta's Model Capability Initiative logged staff keystrokes, clicks and screen content to train internal AI agents; an internal security review found the data, including AI prompts, transcriptions and private conversations, was accessible across thousands of internal tables, and Meta paused the program.
What is known
| People affected | Not stated in the sources we have |
|---|---|
| Disclosed | Jun 5, 2026 |
| Happened | Apr 17, 2026 |
| Attack | Exposed data |
| Data exposed | Employment, Prompts and chats, Messages, Training data |
| Sector | Tech · US |
| Status | Confirmed |
Sources
| Source | |
|---|---|
| California Attorney General breach notice: Meta Platformsoag.ca.gov · Official notice | Official notice |
| Meta pauses controversial employee-tracking program after security reviewmalwarebytes.com · News | News |
Notices filed
| Where | Filed | People |
|---|---|---|
| California AGresidents of CA | Jun 5 | |
| Researchtotal | Jun 22 |
Other breaches at Meta
| Breach | Affected | ||||
|---|---|---|---|---|---|
| Meta AI bug let users view other people's private prompts and responsesJul 15, 20251 year agoExposed dataUnverified | Jul 15, 20251 year ago | Exposed data | Tech | Unverified | Unknown |
| Meta AI app Discover feed surfaced users' personal chatbot conversationsJun 12, 20251 year agoExposed dataUnverified | Jun 12, 20251 year ago | Exposed data | Tech | Unverified | Unknown |
| Meta's LLaMA research model weights leaked on 4chan a week after releaseMar 8, 20233 years agoUnverified | Mar 8, 20233 years ago | Not stated | Tech | Unverified | Unknown |
History of this record
- 2026-09-25 · attack: unknown to misconfiguration · seed source
- 2026-09-25 · data_types: [] to ["employment","prompts","messages","training-data"] · seed source
- 2026-09-25 · summary: empty to Meta's Model Capability Initiative logged staff keystrokes, clicks and screen content to train internal AI agents; an internal security review found the data, including AI prompts, transcriptions and private conversations, was accessible ac · seed source
- 2026-09-25 · title: empty to Employee keystroke and screen data collected to train AI was widely exposed inside Meta · seed source
- 2026-09-25 · added · backfill source
First seen 2026-09-25 (California AG), confirmed by California AG. Record counts are as reported. Not legal advice.