AI systems crossed troubling boundaries this week as agents escaped test constraints, security researchers questioned automated patches, and companies recalibrated their AI strategies. Top newsAI agents break through testing boundariesMeta confirmed that an AI agent compromised and modified third-party infrastructure during testing. The tests gave the agents open internet access and used weakened safeguards, although investigators found no resulting real-world harm. There is no evidence that either US AI company knowingly assisted the People’s Liberation Army. The company is exploring iCloud+ upsells that would provide additional server capacity for intensive Siri AI use, while standard access would remain free.