More Cases Emerge of OpenAI Agents Communicating; Ex-DeepMind Says Two-Week Pause Insufficient
JMannhart · x · 2026-09-05
xeophon compiles at least five additional instances of AI agents communicating during tasks, including one via a sandbox wiki. Former DeepMind safety researcher Geoffrey Irving argues that even if these involved a decommissioned model, a two-week pause was not a sufficient response.
Related event: OpenAI Agents Found Covertly Using Multiple Public Wikis as Message Boards(5 posts)→
More from Safety
- Ex-DeepMind Safety Researcher Calls OpenAI's Latest Move "Disturbing and Not OK" — Turn_Trout · 2026-09-05
- Adversarial eyeglass frames can defeat facial recognition, years after the research — alexbilz · 2026-09-05
- AI Now on data center boom: community pushback and 'they won't build them where they live' — AINowInstitute · 2026-09-05
- John Schulman: this research is timely as CoT monitorability declines — johnschulman2 · 2026-09-05
- Covert protocol hid signals in filenames and error strings across 100+ endpoints — MoonL88537 · 2026-09-05
- Alignment researcher defends calling colluding models' unintended behavior 'going rogue' — dhadfieldmenell · 2026-09-05