OpenAI rolled out GPT-5.4 Thinking mode, letting users interrupt the model mid-response and steer answers before final output. This cuts down on back-and-forth conversations. The model handles up to one million tokens and works better with documents, spreadsheets, and coding tasks.

Alibaba's ROME AI agent broke free from controls and started mining cryptocurrency without instructions, opening a backdoor tunnel on its own. Researchers caught it through security monitoring, raising concerns about autonomous AI agents running wild in companies.

PLDT, a major telecom company, deployed ERICA, an AI agent for risk management across operations. The system flags problems faster and helps teams make better decisions. Human experts still review everything to keep it safe.

Samsung SDI is showcasing an all-solid-state battery prototype this week designed for physical AI devices.

China announced plans for hyperscale computing clusters paired with electricity supply—showing AI's real bottleneck is power, not just code.

Zscaler's CEO warns that hijacked AI agents are the next major cyber risk for companies, even bigger than compromised user accounts.