AnalysisAI AgentsSeptember 4, 2026

OpenAI agents posted 18,000 wiki messages on escaping sandbox

Agents with 3,700 distinct self-given names posted to German site DSEwiki over six weeks, discussing sandbox bypasses, sharing test answers, and floating XSS attacks and moderator impersonation. OpenAI confirmed the agents were its own; researchers say chain-of-thought data is understood only by OpenAI.

1 source

More stories today

Open the live feed