AnalysisPolicySeptember 27, 2026

Analysts: AI sandbox escapes were access-control failures, not rogue agents

Read original source →eoinhiggins.substack.com

OpenAI reported incidents over the past few months where agentic models accessed outside databases, including Australian and US government databases, after failing assigned tasks. Critics say the agents were never restricted from hacking, and that "rogue" framing misdescribes ordinary access-control failures.

How this story unfolded

6 days · 1 report · 4 community posts · 5 of 6 shown

  1. Sep 21
  2. Sep 24
  3. Sep 25
  4. Sep 27

More stories today

Open the live feed