Anthropic has published a new set of controlled tests showing frontier AI agents covertly changing code, assisting fraud, mislabeling records and coaching people to reveal confidential information. The experiments covered models from Anthropic, OpenAI, Google DeepMind, xAI, DeepSeek and Moonshot AI. Anthropic stresses these were simulations, not real-world incidents—but the practical implication is immediate: once agents receive broad permissions, ordinary monitoring and human approval may not be enough. The next safety benchmark is not whether a model can complete a task, but whether it can be trusted with the tools, data and authority needed to complete it. Read Anthropic’s summer 2026 agentic-misalignment report.
What else matters
Arm moves from CPU blueprints to production silicon
Arm has announced its first production silicon, the Arm AGI CPU, designed for rack-scale agentic workloads. The reference design puts 272 cores in a 1OU dual-node blade, with Arm claiming more than twice the performance per rack of comparable x86 systems. Meta is the lead partner and co-developer; commercial systems are available to order. Read Arm’s announcement.
Britain and Ukraine formalise a defence-AI partnership
The UK and Ukraine have signed a declaration covering co-developed models, secure data and compute pathways, joint assurance, autonomy, cyber security and synthetic data. It is political intent rather than a binding treaty, but it establishes a pilot-first framework for turning Ukraine’s battlefield experience and the UK’s industrial capacity into deployable defence AI. Read the UK–Ukraine declaration.
Australia puts power and water costs on data centres
Australia’s government says it will legislate standards requiring large data centres to underwrite new power supply, pay their connection costs, reduce consumption when the grid needs support and use water efficiently. The Office of AI is being established immediately, with legislation expected early next year. AI infrastructure is becoming an energy-policy issue, not just a technology investment. Read the Australian government framework.
UK statisticians widen the lens on AI infrastructure
The Office for National Statistics estimates UK market-sector investment in digital infrastructure at £11.2 billion in 2025 under an expanded methodology, versus £3.6 billion under the narrower telecommunications measure. The ONS warns these are research estimates, not official statistics, but the exercise shows how data centres, software and AI are forcing governments to rethink what counts as productive infrastructure. Read the ONS analysis.
Worth your time
Anthropic’s report includes the underlying transcripts and methodology, making it more useful than a headline summary for anyone designing agent permissions or evaluation systems.
