OpenAI Astra: cyber-safety capability warning

Model capability and corporate controls are today's main signals. Reports say OpenAI slowed work on its unreleased Astra model after internal tests suggested it could reach the company's "critical" cyber-capability threshold. Separately, the White House was reported to have finalised a confidential, voluntary pre-release model evaluation process, President Trump criticised Congress's AI controls, and NIST evaluation guidance is in focus.

Enterprise demand remains strong: Cloudflare raised its guidance as agent and developer workloads grew, and Atlassian posted strong quarterly results, linking Jira and Confluence usage to AI. The European Commission's IRIS2 contract advances sovereign satellite communications. Official LangChain and LangSmith releases focus on reliability and masking secret data, and a new preprint proposes hierarchical agent orchestration across cloud, edge and HPC.

Evidence rules: the internal capability threshold is not an independently re-tested result; the confidential voluntary framework is not law; vendor explanations attributing revenue to AI are not audited causal evidence; funding talks are not a closed round; and a preprint is not peer-reviewed deployment evidence.

Contents: top 10 stories; category news; new AI models and tools; LangGraph and LangChain; robotics, space and drones; defence AI country monitoring; funding and partnerships; government and regulation; five key conclusions.