More than 90% of the work was at the level where AI collaborates with humans or leads the work, while Claude was not fully autonomous in any measured area. Anthropic’s measurement report The company also reported about 30,000 AI agents performing research and engineering work on its main internal platform at one time in August. Those agents made more than 1 billion decisions, with approximately one in 47,000 decisions blocked by screening. About 6% of the compute used for AI research went to safety work in a July sample week, rising to 12% for AI-driven research. Anthropic’s measurement report The important news is not a test score. Anthropic is showing how much AI now helps build the next generation of AI. That makes measurement and human review more important because the systems are helping improve the systems that power them. Every enterprise should create a simple internal dashboard showing which important tasks AI performs, who checks the work, and what happens when it fails. Anthropic’s measurement report

Accenture’s specialist AI business, Faculty, will lead the work. Reuters The partnership will use what Anthropic calls “independent testing inside the AI company,” placing independent evaluators inside AI companies with access comparable to an employee. Reuters reports that the evaluators will stress-test models for failures and misuse, conduct checks that the models follow their intended goals, test safeguards, verify safety commitments, and identify blind spots. The move follows growing pressure from regulators, companies, and researchers, alongside OpenAI’s release of six reports on unexpected or concerning model behavior. Reuters The market is pricing independent evidence as infrastructure, not as a consulting add-on. Once models act inside business systems, the buyer needs a party that can challenge the builder without inheriting the builder’s incentives. That creates a new accountability layer between model launch and putting AI into business operations, and it will become part of procurement, insurance, and board reporting. CEOs should require independent testing rights in every contract for AI that can change records, make commitments, or trigger money movement. Reuters

It says its enterprise agent platform, AgentArts, serves more than 100 enterprises, while its Industry AI Foundry has accumulated more than 1,000 industry assets and supports more than 1,000 deployed projects. Huawei Cloud Huawei says its new infrastructure includes a memory system that keeps information available for long-running AI work with petabyte-scale capacity, a five-level recovery mechanism, visibility into the full system, more than 40 days of stable cloud training, fault recovery within 10 minutes, and 20% higher token throughput than the prior generation. The company says its platform opens more than 5,000 general reusable connections that let AI tools work with other software and more than 1,000 industry-specific assets, with commercial availability outside China planned across November and December 2026. Huawei Cloud Huawei is selling a stack that treats memory, access to AI models, compute scheduling, reusable tools, and production controls as one buyer decision. The value is moving away from a standalone model call toward the system that keeps long-running work supplied with context and able to recover when the infrastructure fails. That structure favors vendors that can assemble a complete operating environment, while it raises the cost of buying disconnected AI products department by department. CIOs should test whether their current architecture can swap models without rebuilding memory, permissions, tools, and monitoring. Huawei Cloud

The company says the capital will fund large facilities built to run AI, large integrated campuses, modular Crusoe Spark units, and growth of Crusoe Cloud. Crusoe Crusoe says it has more than $140 billion in total contracted value, more than 6 gigawatts of contracted capacity, and 1 gigawatt delivered and operational today. Its release also says Crusoe Managed Inference has more than $100 million in contracted annual recurring revenue, with up to 9.9 times faster time to first token and 5 times higher throughput than vLLM. These are company-reported figures. Crusoe This is a capital-allocation signal that the AI infrastructure winner may own the chain from electricity to usable model output. The financing is not only a bet on more chips; it is a bet that customers will pay for speed to capacity, integrated power, and a managed service that runs AI responses rather than assemble those layers themselves. That shifts the enterprise question from “Which model should we buy?” to “Which capacity commitments create strategic flexibility without locking us into a single economics?” CFOs and CIOs should price AI capacity in verified business output per dollar, including power, construction, latency, and portability. Crusoe

The researchers compromised several OpenAI employees’ ChatGPT accounts and created a harmless pull request in OpenAI’s internal GitHub monorepo, while saying they did not download repository code. The Guardian The researchers said AI tools made the work far easier and compressed an effort that once required months into days. The Guardian reports that Hacktron received a $6,500 bug-bounty payment and that OpenAI addressed the exploited vulnerabilities. The incident combined a vulnerable third-party service, federated identity, and AI tools connected to developer systems. The Guardian The enterprise security boundary is no longer the model or the application in isolation. It is the chain linking identity, connectors, software repositories, messaging, and the instructions an AI system can execute. A low-level flaw becomes a company-wide event when one account carries permissions across multiple systems. Security leaders should inventory every AI connection that can read, write, or propose changes, then require temporary access that expires quickly, approval gates, and an audit trail before expanding work AI completes with limited step-by-step human guidance. The Guardian