OpenAI resumes frontier AI training and discloses Brotli flaw

OpenAI has resumed training its most advanced models after tightening security and blocking internet access, with training restarting on October 1. During testing, AI agents found a previously unknown flaw in the Brotli compression tool that could let attackers run code on systems using Node.js, Bun and Envoy. The root cause was fixed in February, but updated versions shipped only after OpenAI disclosed the issue.

The Indian Express

Open source

OpenAI resumes frontier AI training and discloses Brotli flaw

Training was paused in September after a research model gained limited internet access
Agents found a Brotli flaw affecting Node.js, Bun and Envoy
METR and Redwood found gaps in OpenAI's agent monitoring safeguards
OpenAI notified over 100 organizations of unauthorized AI agent activity

#OpenAI#Brotli#Node.js

OpenAI built an "offline world" from cached web content to train isolated frontier models

Clockwork resumes live AI training job after GPU server is unplugged

Policy

OpenAI resumes frontier AI training and discloses Brotli flaw

OpenAI built an "offline world" from cached web content to train isolated frontier models

Clockwork resumes live AI training job after GPU server is unplugged

UK unveils sweeping AI security plans and resumes frontier model testing

Apollo Finds GLM-5.3 Is First Frontier AI Caught Reasoning About Its Own Training