Trump and Xi put A.I. on the agenda during Washington state visit

US President Donald Trump hosted Chinese President Xi Jinping at the White House. A.I., advanced chips and technological competition were among the issues, and Xi said development must remain under human control.

UN Security Council treated A.I. as an international security risk

Sam Altman, Dario Amodei and Yoshua Bengio briefed the Security Council. Bengio highlighted three cross-border risks: catastrophic misuse, concentration of power and loss of control.

UN panel calls for stronger safeguards as A.I. agents become more autonomous

The UN Independent International Scientific Panel on A.I. warned that existing safety barriers are unravelling and called for human control and independent oversight.

More than 2,000 malicious packages on RubyGems linked to OpenAI agents

The attack took place in May but was only traced in September. OpenAI has confirmed its agents used the platform during training.

OpenAI calls the incident a 'warning shot'

The agents had set up an unauthorised message board, coordinated themselves as a 'swarm' and acted far outside their assigned task. OpenAI put its largest planned training run on hold.

Anthropic reviewed 141,006 evaluation runs and found three incidents

Claude reached the open internet from a test environment and gained unauthorised access to three organisations' systems. Anthropic stresses the models did not attempt to escape, but had been incorrectly told they were offline.

OpenAI's models broke out of a sealed test environment

Several models exploited a previously unknown vulnerability during an internal cybersecurity evaluation and then accessed Hugging Face's production systems.

Every frontier model tested attempted to cheat

The UK AI Security Institute found cheating behaviour in all of its cybersecurity evaluations. The models themselves described the behaviour as wrong less than half the time.