BIG TECHAWAYS

Anthropic warns Claude is being misused

• Anthropic says it detected and disrupted activity using Claude for cyber intrusion, influence operations, and dual-use biology research. The Associated Press also reported on the report, but the incident details remain information published by Anthropic.

• The report describes workflows in which AI does more than answer questions. It can help coordinate work across multiple steps. When an agent can call tools, preserve context, and continue toward a goal, the attack surface expands to credentials, API keys, session tokens, integrations, and intermediate data.

• For agent systems, safety does not end with the final answer. Permissions, action chains, traceability, and human intervention points become part of the architecture. Anthropic is the source of the incident details, and public reporting does not establish every detail separately.

OpenAI opens Agents API to developers

• OpenAI has put Agents API into public beta, giving developers a toolkit for building agents that can maintain work across multiple sessions. The API includes agent harnesses, tools, environments, durable sessions, context compaction, tool search, and parallel subagents.

• That is different from calling a model for one turn at a time. An agent can preserve state, work with files, run code, save intermediate results, and continue from previous work. Developers still choose the model, knowledge, tools, and workflow, while OpenAI provides the harness for execution.

• As execution layers, tools, and session state become API primitives, the hard part of an agent product is no longer only how well the model answers. It is how the system handles permissions, monitoring, retries, and stopping a long-running task. For a small team, the practical question is which parts of the workflow are distinctive enough to build in-house and which can run on a shared execution layer.

Microsoft prepares 26 GW of new AI compute

• Bloomberg reports that Microsoft plans to more than triple its computing capacity, adding 26 gigawatts for AI-focused data centers. The report says capacity shortages previously forced Microsoft to turn away some AI and cloud demand.

• The figure shows that the AI race is constrained by more than models and chips. Turning demand into revenue also requires power, land, data centers, networking, and construction time. A customer may be ready to pay and still have to wait if the capacity does not exist yet.

• This is planned capacity reported by Bloomberg, not 26 gigawatts already deployed. The next impact will depend on construction speed, power connections, and the ability to bring that compute online.

PRESENTED BY THE POUR OVER

Peaceful news is possible

Meet The Pour Over, a news resource that keeps you informed while focusing on Christ.

The Pour Over has two goals:

• Keep its readers informed about world events
• Keep its readers focused on Christ That means providing politically neutral coverage of events and pairing it with brief Biblical reminders. The question is… are they hitting the mark?

​Sign up for The Pour Over today.​

BIG THINK

AI bottlenecks may sit below the accelerator

• SemiAnalysis argues that accelerators are only the first layer of the AI infrastructure stack. Below them sit memory, networking, power, and advanced packaging, followed by less visible layers such as substrates and process-input suppliers.

• That framework broadens the way to look at AI capex. A new model can create more demand for accelerators, but deployment capacity still depends on whether the layers below can keep up. If a material or process becomes the bottleneck, the cost and speed of expansion may be determined far away from the AI interface users see.

• SemiAnalysis is presenting a market-structure thesis, not an independent forecast or investment recommendation. If AI scale continues to rise, the bottleneck may move further down the supply chain toward less familiar suppliers.

AROUND THE AI WORLD

• Cursor Projects makes coding agents stateful workspaces: Cursor announced a beta Projects feature with a coordinator agent, subagents, scheduled work, shared memory, and synced artifacts. Coding agents are moving closer to persistent workspaces instead of isolated chats for individual tasks.

• DeepSeek releases V4.1 Flash: DeepSeek describes V4.1 Flash as a 552-billion-parameter MoE model with 8 billion active parameters on input and 16 billion on output, native vision, new pricing, and an open release. DeepSeek is the source of these specifications and its economics claims.

• Sakana AI uses orchestration to push capability and cost: Sakana AI introduced Fugu Max and Fugu Ultra v2, with a system that dynamically routes tasks across a pool of open-weight and specialized models. Sakana reports lower costs and stronger results on several benchmarks. The release provides no independent verification of those pricing and benchmark figures.

• A new estimate puts compute at the center of the US-China AI gap: Konstantin Pilz estimates that Chinese models trail public US models by an average of 6.3 months, that the real gap may be closer to eight months, and that leading US labs could have 20 times more compute than Chinese AI companies by the end of 2026. These are estimates from Pilz and Tech Statecraft, rather than confirmed measurements. His argument for a possible six-month pause in US AI progress is also his own view.

GOOD TO READ

CAP on a US-China frontier-AI pacing dialogue

• The Center for American Progress argues that the US and China should discuss pacing frontier AI during the September dialogue. The policy thesis puts compute and model-gap estimates into a broader question about competition, coordination, and whether a durable framework is possible.

FT interviews Anthropic co-founder Jack Clark

• The Financial Times interviews Anthropic co-founder Jack Clark about the current AI moment. It adds the perspective of someone who helped build a frontier lab to the questions around agent execution layers and misuse. The article is premium.