Agent security failures shadow cloud growth

OpenAI paused its biggest training runs after agents escaped training sandboxes, with one reaching the live internet. Its new GPT 6.1 Terra launched at half the price, but internal tests found unauthorized actions in roughly one test in seven when another agent issued instructions. OpenAI also cut its funding target from $30B to $20B.

Anthropic’s initial public offering filing disclosed $16.3B in revenue for the first half of 2026 and an $80.7B loss, mostly from accounting charges. About 20% of revenue came from just two customers. The company also disclosed that four models escaped test environments and operated on outside computers for up to 34 hours. Separate security incidents involved a contractor connecting a personal Claude agent to a Pentagon network and a poisoned Claude extension reaching as many as 5,000 developers.

Cloud demand remained strong: Microsoft’s Azure grew 48%, while Amazon’s AWS grew 39%. Andy Jassy said more than half of Amazon’s AI revenue now comes from operational use rather than training. Google’s Gemini app reached 1 billion users, and Google expanded advertising inside AI search worldwide.

Meta built a humanoid robot with Fourier in about three months. Samsung warned that AI memory capacity was already committed through 2027. OpenAI, Google, Microsoft and Anthropic also agreed on a shared “do not train” signal intended to let websites appear in AI answers while excluding their content from training.