Claude Security enters beta as Pentagon signs AI deals without Anthropic

Anthropic made Claude Security available in public beta to Claude Enterprise customers on April 30, 2026.¹ The tool scans codebases for vulnerabilities, validates its findings, and proposes patches.¹ It uses Claude Opus 4.7 as its underlying model.¹ Customers can access Claude Security from the Claude.ai sidebar without setting up a custom API integration.²¹

The beta follows an earlier research-preview phase in which the tool was called Claude Code Security.²¹ Rather than examining files in isolation, it can trace data flows and analyze relationships across modules.¹ That can help security teams assess potential vulnerabilities in the context of an application. Proposed patches are not automatically approved: teams still need to review findings and changes before applying them.¹

The US Department of War announced agreements on May 1, 2026, with eight technology companies to deploy their AI models on classified networks. The companies are SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle; Anthropic is not on the list. The announcement gives no specific deployment date.¹

The agreements broaden the group of model providers for highly protected military systems. Anthropic had previously been the only provider of an AI model authorized for classified military operations, according to The New York Times.² Reuters reports that its exclusion is linked to a dispute over safeguards for military use.³ The department describes the arrangements as agreements but does not disclose their value or specify which models will be deployed.¹

Ant Group’s AI lab InclusionAI introduced Ling-2.6-1T on April 30, 2026, as an open model with one trillion parameters.¹ It is designed for reasoning and complex task execution.¹² Its weights are reported to be available on Hugging Face and ModelScope.² The Mixture-of-Experts architecture activates only part of the model for each token, while an adaptive-compute mechanism aims to avoid unnecessary steps.²

The Ling family had already reached this scale in 2025. Ant describes the 2.6 series as an effort to improve the balance between capability and compute use, as well as collaboration among agents.³ Open weights give companies room to test and adapt the model, but its size makes operating costs important to assess.¹² Published benchmark comparisons do not yet establish an advantage in specific business workflows.²

IBM released Granite 4.1 30B as an openly available instruction-tuned model on April 29, 2026. It has 30 billion parameters and is licensed under Apache 2.0.¹ Unsloth subsequently published converted, quantized GGUF files for local inference.²³ On May 6, LM Studio listed the model using a community GGUF build.⁴

The distinction matters: IBM published the base model, while the GGUF files described here come from a third party.¹² Companies can use them to test local inference instead of relying exclusively on a hosted API.² IBM lists summarization, classification, and tool calling among potential uses.¹ Results for the base model should not be assumed to hold for quantized versions, however. Teams should test quality and hardware requirements with the specific build they plan to deploy.

On May 6, TrendForce estimated that nine major cloud providers could spend about $830 billion in capital expenditures in 2026. It identified North American AI data center expansion as a major driver. The figure is an industry forecast, not a joint investment commitment announced by the providers.¹

The scale shows how heavily data centers, servers, and supporting infrastructure feature in spending plans. Individual companies have also raised their forecasts: Alphabet plans $180 billion to $190 billion in 2026 capital expenditures, while Meta projects $125 billion to $145 billion.¹² These budgets cover more than GPUs alone. Comparisons between aggregate figures also depend on which companies are included and whether their forecasts use calendar or fiscal years.¹

Elon Musk testified in a California federal court on April 30, 2026, that xAI had partly used outputs from OpenAI models to train Grok through distillation. Asked whether xAI had done so, he answered, “Partly.” The exchange took place during Musk’s lawsuit against OpenAI and concerned Grok’s training, not the introduction of a new product.¹²

Distillation lets one model learn from another model’s outputs, and Musk described the method as common industry practice.¹² His testimony brings a question for developers and providers into focus: under what conditions can model outputs be used to train a competing system? The testimony alone does not establish that xAI breached OpenAI’s terms of service.¹²