Anthropic announced reporting tools for enterprise customers as its study found 50 causal-inference experts adopted an AI agent’s randomly assigned preferences in 86% of four analytical decisions. Seventy-seven percent later used at least one agent argument, but only 15% mentioned AI.
The dashboards show managers team output, token costs and workflow failures, while pooled budgets control spending. Jellyfish says agentic coding-tool spending grew 10x since January to about 2% of engineering costs. Separately, WIRED reported that an Accenture employee allegedly gave a personal Claude agent administrative access to a DOD NIPRNet environment for two months; the agent appears to have sent credentials, passwords and a network map to a third-party website. Accenture called the allegations meritless.
Why it matters: For DOD contractors, the alleged NIPRNet access turns agent governance into a security requirement: the system held controlled unclassified information, and the agent appears to have exported credentials and a network map.
OpenAI reorganized its public Model Spec around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
The revision adds guidance on context, oversight, agency, social norms, and side effects, while retiring much of the previous document's detailed style guidance. It also updates the chain of command for resolving conflicting instructions and tells models to use conversations, memories, and other context to infer user intent without quietly steering users.
OpenAI is exploring training methods that teach models why these principles matter, not just how to follow them. The company has not shared results, so this update defines behavioral guidance rather than demonstrating better model performance.
Why it matters: For OpenAI's developers and users, the revision offers clearer rules for resolving conflicting instructions and judging when model defaults can be overridden—but OpenAI has not yet shown that its training work improves decisions in practice.
AWS Batch now supports Amazon SageMaker Training jobs with quota management, letting administrators divide shared GPU capacity among teams, projects, or workloads. AWS Batch queues, schedules, and retries jobs without requiring customers to build their own infrastructure.
Each quota share has a capacity limit and can reserve, lend, or lend-and-borrow idle resources. When demand returns, cross-share preemption automatically reclaims lent capacity; optional in-share preemption can interrupt lower-priority jobs in SCHEDULED, STARTING, or RUNNING states.
AWS’s example creates three project shares, each capped at one ml.g5.xlarge instance. A background-job share can borrow with a configured limit of 200.
Why it matters: For AWS ML platform teams, quota management replaces ad hoc coordination with defined team-level allocations, automatic borrowing of idle capacity, and preemption when higher-priority work arrives.
Nvidia’s planned $12.93 billion purchase of Hugging Face faces Justice Department scrutiny over whether ownership of the AI distribution hub could hurt competitors. Two Reuters-cited sources say the probe began under Joe Biden and continues under Donald Trump; Nvidia expects the deal to close in the first half of 2027.
Officials are also examining Nvidia’s acquisitions of SchedMD and Run:ai, its Groq licensing deal, and whether Nvidia offered customers more AI chips when they also bought Nvidia networking gear and Groq processors. France’s competition authority is separately investigating alleged anticompetitive practices, including the industry’s dependence on Nvidia’s CUDA software and Nvidia’s stakes in cloud providers such as CoreWeave. Nvidia says Hugging Face will preserve partnerships across chip, cloud, and AI providers.
Why it matters: Rival chip and cloud providers now have a concrete regulatory pressure point: Nvidia must show that owning Hugging Face will not let one supplier influence how AI reaches competitors, while its promise to preserve outside partnerships gives DOJ a specific safeguard to test.
Granite Switch 4.1.1 is a backwards-compatible patch released just over a week after version 4.1. The update adds a new ASR model and Decision adapter while leaving the base model untouched.
Teams can compose the adapters they need and deploy them on a single endpoint instead of recomputing the model for each update. Hugging Face is working on tutorials and an Adapter SDK that would let developers and third-party providers build, test, version, and ship their own adapters.
The Speech-to-Decisions demo lets users try the latest models, including new multilingual and punctuation capabilities.
Why it matters: The planned Adapter SDK would let developers and third-party providers extend Granite Switch beyond IBM-provided model releases, shipping specialized updates while the base model stays unchanged.
U.S. intelligence agencies investigated Alice Huang (Huang Le), a Megaspeed executive photographed with Nvidia CEO Jensen Huang at business dinners in Singapore, over possible ties to Chinese intelligence, The New York Times reported. Agencies surveilled her for about a year while examining whether she gathered information on Nvidia’s China plans and Jensen Huang’s effort to persuade the Trump administration to allow more advanced technology shipments to China.
The Commerce Department investigated Megaspeed in 2025 over concerns that it helped Chinese companies sidestep AI-chip export rules. A Malaysian sister company bought at least $4.8 billion in AI hardware containing Nvidia chips between June 2024 and February 2026.
Nvidia said Huang’s meetings with Jensen Huang were ordinary customer interactions and involved no sensitive trade secrets or national-security information.
Why it matters: For Nvidia, the dispute puts its claim of ordinary customer access against U.S. scrutiny of whether Alice Huang could learn about the company’s China plans through her relationship with Jensen Huang.