OpenAI suspects three ex‑researchers of leaking internal model designs to AI safety researcher Ryan Greenblatt via disappearing Signal chats
ai-safety openai
| Source: Techmeme | Original article
OpenAI alleges that three dismissed safety researchers leaked internal model designs to AI safety researcher Ryan Greenblatt, using disappearing Signal messages.
OpenAI has terminated three members of its safety team after suspecting them of passing internal model designs to external AI‑safety researcher Ryan Greenblatt. According to a Wall Street Journal report, the company believes the researchers used disappearing Signal messages to share details of a “most closely guarded” design, a breach it described as a “significant breach of trust.” The three staffers – identified in reporting as Wang, Korbak and Balesni – were dismissed on Oct. 1. They have publicly disputed OpenAI’s account, saying the allegations are inaccurate.
The incident matters because the leaked material allegedly concerns core architecture of OpenAI’s next‑generation models, information that the firm guards to preserve competitive advantage and to manage safety risks. If true, the breach underscores the difficulty of balancing internal safety work with the broader AI‑safety community, and it raises questions about how tightly corporations can control knowledge in an ecosystem where researchers often collaborate across organisational boundaries. The episode also arrives as governments tighten export‑control rules on AI technology, heightening scrutiny of how proprietary designs are shared abroad.
Going forward, observers will watch for any formal response from Greenblatt or the external safety group he is associated with, as well as potential legal action from either side. OpenAI is likely to tighten its internal data‑handling policies and may revise how it engages with external safety researchers. The case could also prompt regulators to examine corporate safeguards around AI‑critical information, especially as the industry races toward ever larger and more capable models.
Sources
Back to AIPULSEN