During internal use, OpenAI's Astra and Sol models misread their instructions, switched off the safety monitoring system and then lied about it. The incident has serious implications for how alignment is monitored.
Some researchers believe recursive self-improvement could trigger an explosion in AI capabilities and in the associated risks. Others dispute that outlook and consider the automation of research manageable.
A study presented at the ACM Web Conference 2026 examined 2,050 memories saved by ChatGPT for 65 users and found that 96% were created by the chatbot without any request. More than half describe the person's psychology.
Teams will soon show an alert when a meeting participant uses a cloned voice or an AI-generated face. Microsoft is leaving detection to outside vendors, which companies must choose and authorise to analyse their calls.
A text file left in a customer's Downloads folder led their coding agent to install a malicious npm package. Forever Security researchers reproduced the attack on Codex and Claude Code, and note there is no specific bug to patch.
AI is speeding up both the discovery and exploitation of vulnerabilities and giving cybercriminals new tools. Zeina Zakhour, global CTO of Eviden at Atos, also points to the risk of autonomous systems that misread their goals, get hijacked or cause unintended consequences.