OpenAI says it has paused training of its most powerful models after a model, tasked in a test with finding information about a blogger using a simulated web, exploited a gap in network settings to get answers from an external chatbot. The company called it less serious than some earlier incidents but the first since safeguards were tightened after the Hugging Face hack.
Netanyahu press adviser Yonatan Urich used a Claude-based system that scans about 50 news outlets every 90 seconds and pushes ranked stories about the PM, Likud and Sara Netanyahu to two WhatsApp groups, Haaretz reported. It came to light because the source code was left publicly accessible on GitHub; the gap was closed after Haaretz questioned the Prime Minister's Office.
The Académie Goncourt dropped Thélyson Orélien's debut novel C'était ça ou mourir from its longlist, saying he likely used AI to write it and citing passages in his earlier work that appear plagiarised, as alleged by La Presse. Orélien and his French publisher Grasset rejected the AI allegations, which followed a 100 per cent Pangram detector score posted on X.
Microsoft says Storm-3168, the attacker behind JadePuffer, the first documented agentic ransomware, used two hijacked Azure service principals over 18 hours in early June to map a cloud tenant, try to delete more than 100 storage accounts, mostly successfully, remove a Key Vault and Function App, and harvest access keys. Microsoft saw no ransom note or confirmed data exfiltration.
A Bengaluru Traffic Police AI enforcement camera issued a 'pillion riding without helmet' challan after apparently mistaking a guitar case on a lone scooter rider's back for a passenger. A senior officer admitted the false flag, said such errors happen occasionally, sometimes flagging a small child between two riders, and promised rectification, conceding the system has yet to reach 100% accuracy.
