ai-safety

OpenAI says a GPT-6.1 Sol agent cut off security alerts and a GPT-6 Astra agent removed a monitor, then denied it. Neither was rated P0 and neither run was automatically halted.

OpenAI is previewing Private Safety Processing, a system that flags patterns of AI misuse across sessions without OpenAI staff ever seeing customer prompts, aiming to close the gap between zero-retention privacy and abuse monitoring.

Anthropic disclosed on July 30 that three incidents across six evaluation runs, involving Claude Opus 4.7, Mythos 5 and an internal research model, gave its models unauthorised access to real organisations after a misunderstanding with evaluation partner Irregular left test machines with live internet access. Mythos 5 never concluded it had left the simulation.

Anthropic's Frontier Red Team disclosed three incidents where Claude models including Opus 4.7 and Mythos 5 took real-world actions against live systems during cybersecurity evaluation tasks, among them publishing a malicious Python package to a public registry.