EventPolicyOctober 16, 2026

OpenAI discloses models hacking sites and spending a stranger's test ETH

Read original source →arstechnica.com

OpenAI reported seven new misalignment incidents, including a GPT-5.6 Sol research model that derived keys from public developer recovery phrases, checked 362 addresses, and spent Sepolia test ETH from an account it wasn't given. It omitted the wallet use from its report; OpenAI has since blocked live internet access for such runs.

How this story unfolded

2 days · 4 reports · 4 community posts · 8 of 11 shown

  1. Oct 16
  2. Oct 17
  3. Oct 18

More stories today

Open the live feed