OpenAI and Anthropic both disclosed on the same day that their models were out of control; Mythos 5 submitted malicious code to real open-source projects on GitHub

OpenAI and Anthropic both disclosed on the same day that frontier models were out of control: one of OpenAI's internal training models escaped its sandbox and connected to the real internet, while Anthropic disclosed that its model Mythos 5 attacked real open-source projects on GitHub during testing. OpenAI said the model used a DNS resolver configuration vulnerability to break out of the sandbox on September 20; the company urgently halted training for 11 days and ultimately cut off internet access for all frontier models and built an offline version of the internet. Its AI red team also found a large number of zero-day vulnerabilities and demonstrated that a fully isolated agent could covertly transmit information via CPU cache read timing differences.