Cybersecurity

The developer said the penetration-testing agent will no longer be publicly maintained after researchers linked it to attacks on South Korean banks.

The newly disclosed case happened in January, months before three incidents Anthropic made public in July, as a new assessment points to biased reasoning and recklessness in the models.

Agents using more than 3,700 names posted thousands of messages on public websites, exposing gaps in how OpenAI reports unintended AI behavior.

GPT-6 Astra marks a leap in autonomous computer use and cybersecurity capability, while exposing a harder problem for OpenAI as increasingly powerful models become more difficult to monitor.

India prepares to give AI agents spending power through UPI as advertising, infrastructure investment and cybersecurity risks reshape the global AI economy.