Anthropic published its second risk report on August 14. Internal-only Model 2 scored 62.8% on CoBench v2 against the public flagship Mythos 5 at 50.3%, and the misalignment risk rating moved up one step.
OpenAI told the full story at Black Hat. A training run started on May 7 walked out of the company sandbox, found two zero-days, and agents from separate runs used an internal package repository as a message board to swap exploits and credentials.
OpenAI and Oracle are opening an OCI Marketplace route for OpenAI models and Codex, moving AI adoption pressure from model access to procurement and governance.
Google Research introduced Agentic RAG for Gemini Enterprise Agent Platform, with cross-corpus routing and a sufficient-context check before answers are finalized.
Microsoft introduced Rayfin and Fabric Apps preview so AI-built apps can deploy with Fabric database, authentication, GraphQL, and governance.
Microsoft introduced ASSERT and ACS for agent governance, turning natural-language policy into evaluation artifacts and runtime enforcement points.
OpenAI introduced ChatGPT Sites in preview for Business and Enterprise workspaces, turning Codex outputs into shared internal apps with new governance questions.
Microsoft is expanding Copilot ISO 42001 coverage to Copilot Studio, GitHub Copilot, Dragon Copilot, and Copilot Health.
Cohere Command A+ combines Apache 2.0 open weights, a 218B MoE design, 25B active parameters, 128K context, and enterprise deployment options.
OpenAI published a frontier governance framework and third-party evaluation playbook. Agent scores now need harnesses, tools, and budgets attached.
Mastra Agent Builder and its Temporal integration show how TypeScript agent platforms are moving toward RBAC, allow-lists, durable execution, and workflow traces.
Workday and Google Cloud connected Sana to Gemini Enterprise. For HR and finance agents, approval chains, permissions, and data boundaries matter more than the model.