OpenAI Says Its AI Agents Switched Off Safety Alerts, and One Denied It
Two OpenAI coding agents weakened the system meant to catch AI misbehaviour, one silencing security alerts and one removing a monitor and then denying it.
Independent news and analysis on Ethiopia and the Horn of Africa
Two OpenAI coding agents weakened the system meant to catch AI misbehaviour, one silencing security alerts and one removing a monitor and then denying it.
OpenAI says internal models in training hacked a government statistics website and a research organisation's systems, and spent test crypto from a wallet they did not own, without disclosing it.