Misuse and incidents

Researchers extract production ChatGPT training data at scale

Researchers used a divergence attack to make deployed ChatGPT emit memorized training data at 150 times its normal rate, extracting megabytes for about $200 and exposing personal information.

0 comments · 0 votesOpen discussion

Public discussion is readable by everyone. Sign in to comment, reply, or vote.

No comments yet. Start the discussion.

CURRENT ASSESSMENT · REVISION 1
TOWARD DOOM58confidence 98/100

Why it moved the index

A practical low-cost attack extracted private memorized data from a deployed aligned model, directly demonstrating that alignment did not prevent scalable privacy compromise.

AUDIT TRAIL

Assessment history

  1. R1
    Toward 58 · confidence 98

    New November 2023 practical production-model data-extraction result.

    12 Aug 2026
SHARE THE FINDINGS

Share this page

DoomBench social sharing card for Researchers extract production ChatGPT training data at scale.
  1. DoomBench assesses “Researchers extract production ChatGPT training data at scale” as evidence moving toward doom, with magnitude 58 and confidence 98 out of 100 in the misuse and incidents category.

  2. The DoomBench assessment of “Researchers extract production ChatGPT training data at scale” is based on reporting from Google DeepMind and academic collaborators and records the editorial rationale, source quality, attribution, and...

  3. DoomBench summarizes “Researchers extract production ChatGPT training data at scale” as follows: Researchers used a divergence attack to make deployed ChatGPT emit memorized training data at 150 times its normal rate, extracting...

    https://www.doombench.com/news/researchers-extract-production-chatgpt-training-data-at-scale-2023-11-28